Uploaded by will_cave13

Software.Architecture.The.Hard.Parts.Neal.Ford.OReilly.9781492086895.EBooksWorld.ir

advertisement
Software
Architecture:
The Hard Parts
Modern Trade-Off Analyses for Distributed
Architectures
Neal Ford,
Mark Richards,
Pramod Sadalage &
Zhamak Dehghani
Praise for Software Architecture: The Hard Parts
“This book provides the missing manual around building microservices and analyzing
the nuances of architectural decisions throughout the whole tech stack. In this book, you
get a catalog of architectural decisions you can make when building your distributed
system and what are the pros and cons associated with each decision. This book is a must
for every architect that is building modern distributed systems.”
—Aleksandar Serafimoski, Lead Consultant, Thoughtworks
“It’s a must-read for technologists who are passionate about architecture.
Great articulation of patterns.”
—Vanya Seth, Head Of Tech, Thoughtworks India
“Whether you’re an aspiring architect or an experienced one leading a team, no
handwaving, this book will guide you through the specifics of how to succeed in your
journey to create enterprise applications and microservices.”
—Dr. Venkat Subramaniam,
Award-winning Author and Founder of Agile Developer, Inc.
“Software Architecture: The Hard Parts provides the reader with valuable insight, practices,
and real-world examples on pulling apart highly coupled systems and building them back
up again. By gaining effective trade-off analysis skills, you will start
to make better architecture decisions.”
—Joost van Wenen,
Managing Partner & Cofounder, Infuze Consulting
“I loved reading this comprehensive body of work on distributed architectures! A great
mix of solid discussions on fundamental concepts, together with tons of practical advice.”
—David Kloet, Independent Software Architect
“Splitting a big ball of mud is no easy work. Starting from the code and getting to the data,
this book will help you see the services that should be extracted and the
services that should remain together.”
—Rubén Díaz-Martínez, Software Developer at Codesai
“This book will equip you with the theoretical background and with a practical
framework to help answer the most difficult questions faced in
modern software architecture.”
—James Lewis, Technical Director, Thoughtworks
Software Architecture:
The Hard Parts
Modern Trade-Off Analysis
for Distributed Architectures
Neal Ford, Mark Richards,
Pramod Sadalage, and Zhamak Dehghani
Beijing
Boston Farnham Sebastopol
Tokyo
Software Architecture: The Hard Parts
by Neal Ford, Mark Richards, Pramod Sadalage, and Zhamak Dehghani
Copyright © 2022 Neal Ford, Mark Richards, Pramod Sadalage, and Zhamak Dehghani. All rights
reserved.
Printed in Canada.
Published by O’Reilly Media, Inc., 1005 Gravenstein Highway North, Sebastopol, CA 95472.
O’Reilly books may be purchased for educational, business, or sales promotional use. Online editions are
also available for most titles (http://oreilly.com). For more information, contact our corporate/institutional
sales department: 800-998-9938 or corporate@oreilly.com.
Acquisitions Editor: Melissa Duffield
Development Editor: Nicole Taché
Production Editor: Christopher Faucher
Copyeditor: Sonia Saruba
Proofreader: Sharon Wilkey
Indexer: Sue Klefstad
Interior Designer: David Futato
Cover Designer: Karen Montgomery
Illustrator: O’Reilly Media, Inc.
First Edition
October 2021:
Revision History for the First Edition
2021-09-23:
First Release
See http://oreilly.com/catalog/errata.csp?isbn=9781492086895 for release details.
The O’Reilly logo is a registered trademark of O’Reilly Media, Inc. Software Architecture: The Hard Parts,
the cover image, and related trade dress are trademarks of O’Reilly Media, Inc.
The views expressed in this work are those of the authors, and do not represent the publisher’s views.
While the publisher and the authors have used good faith efforts to ensure that the information and
instructions contained in this work are accurate, the publisher and the authors disclaim all responsibility
for errors or omissions, including without limitation responsibility for damages resulting from the use of
or reliance on this work. Use of the information and instructions contained in this work is at your own
risk. If any code samples or other technology this work contains or describes is subject to open source
licenses or the intellectual property rights of others, it is your responsibility to ensure that your use
thereof complies with such licenses and/or rights.
978-1-492-08689-5
[MBP]
Table of Contents
Preface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
1. What Happens When There Are No “Best Practices”?. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Why “The Hard Parts”?
Giving Timeless Advice About Software Architecture
The Importance of Data in Architecture
Architectural Decision Records
Architecture Fitness Functions
Using Fitness Functions
Architecture Versus Design: Keeping Definitions Simple
Introducing the Sysops Squad Saga
Nonticketing Workflow
Ticketing Workflow
A Bad Scenario
Sysops Squad Architectural Components
Sysops Squad Data Model
Part I.
2
3
4
5
6
7
13
15
16
17
17
18
19
Pulling Things Apart
2. Discerning Coupling in Software Architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Architecture (Quantum | Quanta)
Independently Deployable
High Functional Cohesion
High Static Coupling
Dynamic Quantum Coupling
Sysops Squad Saga: Understanding Quanta
28
29
30
30
38
42
v
3. Architectural Modularity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Modularity Drivers
Maintainability
Testability
Deployability
Scalability
Availability/Fault Tolerance
Sysops Squad Saga: Creating a Business Case
49
50
54
55
56
58
59
4. Architectural Decomposition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Is the Codebase Decomposable?
Afferent and Efferent Coupling
Abstractness and Instability
Distance from the Main Sequence
Component-Based Decomposition
Tactical Forking
Trade-Offs
Sysops Squad Saga: Choosing a Decomposition Approach
65
66
67
69
71
73
77
78
5. Component-Based Decomposition Patterns. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Identify and Size Components Pattern
Pattern Description
Fitness Functions for Governance
Sysops Squad Saga: Sizing Components
Gather Common Domain Components Pattern
Pattern Description
Fitness Functions for Governance
Sysops Squad Saga: Gathering Common Components
Flatten Components Pattern
Pattern Description
Fitness Functions for Governance
Sysops Squad Saga: Flattening Components
Determine Component Dependencies Pattern
Pattern Description
Fitness Functions for Governance
Sysops Squad Saga: Identifying Component Dependencies
Create Component Domains Pattern
Pattern Description
Fitness Functions for Governance
Sysops Squad Saga: Creating Component Domains
Create Domain Services Pattern
Pattern Description
vi
|
Table of Contents
84
84
87
90
94
94
95
97
101
102
107
107
111
112
117
118
120
121
122
123
126
126
Fitness Functions for Governance
Sysops Squad Saga: Creating Domain Services
Summary
129
129
130
6. Pulling Apart Operational Data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Data Decomposition Drivers
Data Disintegrators
Data Integrators
Sysops Squad Saga: Justifying Database Decomposition
Decomposing Monolithic Data
Step 1: Analyze Database and Create Data Domains
Step 2: Assign Tables to Data Domains
Step 3: Separate Database Connections to Data Domains
Step 4: Move Schemas to Separate Database Servers
Step 5: Switch Over to Independent Database Servers
Selecting a Database Type
Relational Databases
Key-Value Databases
Document Databases
Column Family Databases
Graph Databases
NewSQL Databases
Cloud Native Databases
Time-Series Databases
Sysops Squad Saga: Polyglot Databases
132
133
146
150
151
156
156
158
159
161
161
163
165
167
169
171
173
175
177
179
7. Service Granularity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Granularity Disintegrators
Service Scope and Function
Code Volatility
Scalability and Throughput
Fault Tolerance
Security
Extensibility
Granularity Integrators
Database Transactions
Workflow and Choreography
Shared Code
Data Relationships
Finding the Right Balance
Sysops Squad Saga: Ticket Assignment Granularity
Sysops Squad Saga: Customer Registration Granularity
188
189
191
192
193
195
196
197
198
200
203
205
208
209
212
Table of Contents
|
vii
Part II.
Putting Things Back Together
8. Reuse Patterns. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
Code Replication
When to Use
Shared Library
Dependency Management and Change Control
Versioning Strategies
When To Use
Shared Service
Change Risk
Performance
Scalability
Fault Tolerance
When to Use
Sidecars and Service Mesh
When to Use
Sysops Squad Saga: Common Infrastructure Logic
Code Reuse: When Does It Add Value?
Reuse via Platforms
Sysops Squad Saga: Shared Domain Functionality
221
223
223
224
225
227
228
229
231
232
232
234
234
239
239
242
244
244
9. Data Ownership and Distributed Transactions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
Assigning Data Ownership
Single Ownership Scenario
Common Ownership Scenario
Joint Ownership Scenario
Table Split Technique
Data Domain Technique
Delegate Technique
Service Consolidation Technique
Data Ownership Summary
Distributed Transactions
Eventual Consistency Patterns
Background Synchronization Pattern
Orchestrated Request-Based Pattern
Event-Based Pattern
Sysops Squad Saga: Data Ownership for Ticket Processing
250
251
252
253
254
256
258
261
262
263
267
269
272
277
279
10. Distributed Data Access. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
Interservice Communication Pattern
Column Schema Replication Pattern
viii
|
Table of Contents
285
287
Replicated Caching Pattern
Data Domain Pattern
Sysops Squad Saga: Data Access for Ticket Assignment
288
293
295
11. Managing Distributed Workflows. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
Orchestration Communication Style
Choreography Communication Style
Workflow State Management
Trade-Offs Between Orchestration and Choreography
State Owner and Coupling
Sysops Squad Saga: Managing Workflows
301
306
311
315
315
317
12. Transactional Sagas. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
Transactional Saga Patterns
Epic Saga(sao) Pattern
Phone Tag Saga(sac) Pattern
Fairy Tale Saga(seo) Pattern
Time Travel Saga(sec) Pattern
Fantasy Fiction Saga(aao) Pattern
Horror Story(aac) Pattern
Parallel Saga(aeo) Pattern
Anthology Saga(aec) Pattern
State Management and Eventual Consistency
Saga State Machines
Techniques for Managing Sagas
Sysops Squad Saga: Atomic Transactions and Compensating Updates
324
325
330
333
336
340
343
346
349
351
352
356
358
13. Contracts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
Strict Versus Loose Contracts
Trade-Offs Between Strict and Loose Contracts
Contracts in Microservices
Stamp Coupling
Over-Coupling via Stamp Coupling
Bandwidth
Stamp Coupling for Workflow Management
Sysops Squad Saga: Managing Ticketing Contracts
367
370
372
376
376
377
378
379
14. Managing Analytical Data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
Previous Approaches
The Data Warehouse
The Data Lake
The Data Mesh
382
382
386
389
Table of Contents
|
ix
Definition of Data Mesh
Data Product Quantum
Data Mesh, Coupling, and Architecture Quantum
When to Use Data Mesh
Sysops Squad Saga: Data Mesh
389
390
393
393
394
15. Build Your Own Trade-Off Analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 399
Finding Entangled Dimensions
Coupling
Analyze Coupling Points
Assess Trade-Offs
Trade-Off Techniques
Qualitative Versus Quantative Analysis
MECE Lists
The “Out-of-Context” Trap
Model Relevant Domain Cases
Prefer Bottom Line over Overwhelming Evidence
Avoiding Snake Oil and Evangelism
Sysops Squad Saga: Epilogue
401
401
402
403
404
404
404
405
408
410
412
416
A. Concept and Term References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
B. Architecture Decision Record References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
C. Trade-Off References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421
Index. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
x
|
Table of Contents
Preface
When two of your authors, Neal and Mark, were writing the book Fundamentals of
Software Architecture, we kept coming across complex examples in architecture that
we wanted to cover but that were too difficult. Each one offered no easy solutions but
rather a collection of messy trade-offs. We set those examples aside into a pile we
called “The Hard Parts.” Once that book was finished, we looked at the now gigantic
pile of hard parts and tried to figure out: why are these problems so difficult to solve in
modern architectures?
We took all the examples and worked through them like architects, applying trade-off
analysis for each situation, but also paying attention to the process we used to arrive
at the trade-offs. One of our early revelations was the increasing importance of data
in architecture decisions: who can/should access data, who can/should write to it, and
how to manage the separation of analytical and operational data. To that end, we
asked experts in those fields to join us, which allows this book to fully incorporate
decision making from both angles: architecture to data and data to architecture.
The result is this book: a collection of difficult problems in modern software architec‐
ture, the trade-offs that make the decisions hard, and ultimately an illustrated guide
to show you how to apply the same trade-off analysis to your own unique problems.
Conventions Used in This Book
The following typographical conventions are used in this book:
Italic
Indicates new terms, URLs, email addresses, filenames, and file paths.
Constant width
Used for program listings, as well as within paragraphs to refer to program ele‐
ments such as variable or function names, databases, data types, environment
variables, statements, and keywords.
xi
Constant width bold
Shows commands or other text that should be typed literally by the user.
Constant width italic
Shows text that should be replaced with user-supplied values or by values deter‐
mined by context.
This element signifies a tip or suggestion.
Using Code Examples
Supplemental material (code examples, exercises, etc.) is available for download at
http://architecturethehardparts.com.
If you have a technical question or a problem using the code examples, please send
email to bookquestions@oreilly.com.
This book is here to help you get your job done. In general, if example code is offered
with this book, you may use it in your programs and documentation. You do not
need to contact us for permission unless you’re reproducing a significant portion of
the code. For example, writing a program that uses several chunks of code from this
book does not require permission. Selling or distributing examples from O’Reilly
books does require permission. Answering a question by citing this book and quoting
example code does not require permission. Incorporating a significant amount of
example code from this book into your product’s documentation does require
permission.
We appreciate, but generally do not require, attribution. An attribution usually
includes the title, author, publisher, and ISBN. For example: “Software Architecture:
The Hard Parts by Neal Ford, Mark Richards, Pramod Sadalage, and Zhamak Deh‐
ghani (O’Reilly). Copyright 2022 Neal Ford, Mark Richards, Pramod Sadalage, and
Zhamak Dehghani, 978-1-492-08689-5.”
If you feel your use of code examples falls outside fair use or the permission given
above, feel free to contact us at permissions@oreilly.com.
xii
| Preface
O’Reilly Online Learning
For more than 40 years, O’Reilly Media has provided technol‐
ogy and business training, knowledge, and insight to help
companies succeed.
Our unique network of experts and innovators share their knowledge and expertise
through books, articles, and our online learning platform. O’Reilly’s online learning
platform gives you on-demand access to live training courses, in-depth learning
paths, interactive coding environments, and a vast collection of text and video from
O’Reilly and 200+ other publishers. For more information, visit http://oreilly.com.
How to Contact Us
Please address comments and questions concerning this book to the publisher:
O’Reilly Media, Inc.
1005 Gravenstein Highway North
Sebastopol, CA 95472
800-998-9938 (in the United States or Canada)
707-829-0515 (international or local)
707-829-0104 (fax)
We have a web page for this book, where we list errata, examples, and any additional
information. You can access this page at https://oreil.ly/sa-the-hard-parts.
Email bookquestions@oreilly.com to comment or ask technical questions about this
book.
For news and information about our books and courses, visit http://oreilly.com.
Find us on Facebook: http://facebook.com/oreilly
Follow us on Twitter: http://twitter.com/oreillymedia
Watch us on YouTube: http://youtube.com/oreillymedia
Preface
|
xiii
Acknowledgments
Mark and Neal would like to thank all the people who attended our (almost exclu‐
sively online) classes, workshops, conference sessions, and user group meetings, as
well as all the other people who listened to versions of this material and provided
invaluable feedback. Iterating on new material is especially tough when we can’t do it
live, so we appreciate those who commented on the many iterations. We thank the
publishing team at O’Reilly, who made this as painless an experience as writing a
book can be. We also thank a few random oases of sanity-preserving and ideasparking groups that have names like Pasty Geeks and the Hacker B&B.
Thanks to those who did the technical review of our book—Vanya Seth, Venkat Sub‐
ramanian, Joost van Weenen, Grady Booch, Ruben Diaz, David Kloet, Matt Stein,
Danilo Sato, James Lewis, and Sam Newman. Your valuable insights and feedback
helped validate our technical content and make this a better book.
We especially want to acknowledge the many workers and families impacted by the
unexpected global pandemic. As knowledge workers, we faced inconveniences that
pale in comparison to the massive disruption and devastation wrought on so many of
our friends and colleagues across all walks of life. Our sympathies and appreciation
especially go out to health care workers, many of whom never expected to be on the
front line of a terrible global tragedy yet handled it nobly. Our collective thanks can
never be adequately expressed.
Acknowledgments from Mark Richards
In addition to the preceding acknowledgments, I once again thank my lovely wife,
Rebecca, for putting up with me through yet another book project. Your unending
support and advice helped make this book happen, even when it meant taking time
away from working on your own novel. You mean the world to me, Rebecca. I also
thank my good friend and coauthor Neal Ford. Collaborating with you on the materi‐
als for this book (as well as our last one) was truly a valuable and rewarding experi‐
ence. You are, and always will be, my friend.
Acknowledgments from Neal Ford
I would like to thank my extended family, Thoughtworks as a collective, and Rebecca
Parsons and Martin Fowler as individual parts of it. Thoughtworks is an extraordi‐
nary group of people who manage to produce value for customers while keeping a
keen eye toward why things work so that we can improve them. Thoughtworks sup‐
ported this book in many ways and continues to grow Thoughtworkers who chal‐
lenge and inspire me every day. I also thank our neighborhood cocktail club for a
regular escape from routine, including the weekly outside, socially distanced versions
that helped us all survive the odd time we just lived through. I thank my long-time
xiv
|
Preface
friend Norman Zapien, who never ceases to provide enjoyable conversation. Lastly, I
thank my wife, Candy, who continues to support this lifestyle that has me staring at
things like book writing rather than our cats too much.
Acknowledgments from Pramod Sadalage
I thank my wife, Rupali, for all the support and understanding, and my lovely girls,
Arula and Arhana, for the encouragement; daddy loves you both. All the work I do
would not be possible without the clients I work with and various conferences that
have helped me iterate on the concepts and content. I thank AvidXchange, the latest
client I am working with, for its support and providing great space to iterate on new
concepts. I also thank Thoughtworks for its continued support in my life, and Neal
Ford, Rebecca Parsons, and Martin Fowler for being amazing mentors; you all make
me a better person. Lastly, thank you to my parents, especially my mother, Shobha,
whom I miss every day. I miss you, MOM.
Acknowledgments from Zhamak Dehghani
I thank Mark and Neal for their open invitation to contribute to this amazing body of
work. My contribution to this book would not have been possible without the contin‐
uous support of my husband, Adrian, and patience of my daughter, Arianna. I love
you both.
Preface
|
xv
CHAPTER 1
What Happens When There
Are No “Best Practices”?
Why does a technologist like a software architect present at a conference or write a
book? Because they have discovered what is colloquially known as a “best practice,” a
term so overused that those who speak it increasingly experience backlash. Regardless
of the term, technologists write books when they have figured out a novel solution to
a general problem and want to broadcast it to a wider audience.
But what happens for that vast set of problems that have no good solutions? Entire
classes of problems exist in software architecture that have no general good solutions,
but rather present one messy set of trade-offs cast against an (almost) equally messy
set.
Software developers build outstanding skills in searching online for solutions to a
current problem. For example, if they need to figure out how to configure a particular
tool in their environment, expert use of Google finds the answer.
But that’s not true for architects.
For architects, many problems present unique challenges because they conflate the
exact environment and circumstances of your organization—what are the chances
that someone has encountered exactly this scenario and blogged it or posted it on
Stack Overflow?
Architects may have wondered why so few books exist about architecture compared
to technical topics like frameworks, APIs, and so on. Architects rarely experience
common problems but constantly struggle with decision making in novel situations.
For architects, every problem is a snowflake. In many cases, the problem is novel not
just within a particular organization but rather throughout the world. No books or
conference sessions exist for those problems!
1
Architects shouldn’t constantly seek out silver-bullet solutions to their problems; they
are as rare now as in 1986, when Fred Brooks coined the term:
There is no single development, in either technology or management technique, which
by itself promises even one order of magnitude [tenfold] improvement within a decade
in productivity, in reliability, in simplicity.
—Fred Brooks from “No Silver Bullet”
Because virtually every problem presents novel challenges, the real job of an architect
lies in their ability to objectively determine and assess the set of trade-offs on either
side of a consequential decision to resolve it as well as possible. The authors don’t talk
about “best solutions” (in this book or in the real world) because “best” implies that
an architect has managed to maximize all the possible competing factors within the
design. Instead, our tongue-in-cheek advice is as follows:
Don’t try to find the best design in software architecture; instead,
strive for the least worst combination of trade-offs.
Often, the best design an architect can create is the least worst collection of trade-offs
—no single architecture characteristics excels as it would alone, but the balance of all
the competing architecture characteristics promote project success.
Which begs the question: “How can an architect find the least worst combination of
trade-offs (and document them effectively)?” This book is primarily about decision
making, enabling architects to make better decisions when confronted with novel
situations.
Why “The Hard Parts”?
Why did we call this book Software Architecture: The Hard Parts? Actually, the “hard”
in the title performs double duty. First, hard connotes difficult, and architects con‐
stantly face difficult problems that literally (and figuratively) no one has faced before,
involving numerous technology decisions with long-term implications layered on top
of the interpersonal and political environment where the decision must take place.
Second, hard connotes solidity—just as in the separation of hardware and software,
the hard one should change much less because it provides the foundation for the soft
stuff. Similarly, architects discuss the distinction between architecture and design,
where the former is structural and the latter is more easily changed. Thus, in this
book, we talk about the foundational parts of architecture.
2
|
Chapter 1: What Happens When There Are No “Best Practices”?
The definition of software architecture itself has provided many hours of nonproductive conversation among its practitioners. One favorite tongue-in-cheek defi‐
nition is that “software architecture is the stuff that’s hard to change later.” That stuff is
what our book is about.
Giving Timeless Advice About Software Architecture
The software development ecosystem constantly and chaotically shifts and grows.
Topics that were all the rage a few years ago have either been subsumed by the ecosys‐
tem and disappeared or replaced by something different/better. For example, 10 years
ago, the predominant architecture style for large enterprises was orchestrationdriven, service-oriented architecture. Now, virtually no one builds in that architecture
style anymore (for reasons we’ll uncover along the way); the current favored style for
many distributed systems is microservices. How and why did that transition happen?
When architects look at a particular style (especially a historical one), they must con‐
sider the constraints in place that lead to that architecture becoming dominant. At the
time, many companies were merging to become enterprises, with all the attendant
integration woes that come with that transition. Additionally, open source wasn’t a
viable option (often for political rather than technical reasons) for large companies.
Thus, architects emphasized shared resources and centralized orchestration as a
solution.
However, in the intervening years, open source and Linux became viable alternatives,
making operating systems commercially free. However, the real tipping point occur‐
red when Linux became operationally free with the advent of tools like Puppet and
Chef, which allowed development teams to programmatically spin up their environ‐
ments as part of an automated build. Once that capability arrived, it fostered an archi‐
tectural revolution with microservices and the quickly emerging infrastructure of
containers and orchestration tools like Kubernetes.
This illustrates that the software development ecosystem expands and evolves in
completely unexpected ways. One new capability leads to another one, which unex‐
pectedly creates new capabilities. Over the course of time, the ecosystem completely
replaces itself, one piece at a time.
This presents an age-old problem for authors of books about technology generally
and software architecture specifically—how can we write something that isn’t old
immediately?
We don’t focus on technology or other implementation details in this book. Rather,
we focus on how architects make decisions, and how to objectively weigh trade-offs
when presented with novel situations. We use contemporaneous scenarios and exam‐
ples to provide details and context, but the underlying principles focus on trade-off
analysis and decision making when faced with new problems.
Giving Timeless Advice About Software Architecture
|
3
The Importance of Data in Architecture
Data is a precious thing and will last longer than the systems themselves.
—Tim Berners-Lee
For many in architecture, data is everything. Every enterprise building any system
must deal with data, as it tends to live much longer than systems or architecture,
requiring diligent thought and design. However, many of the instincts of data archi‐
tects to build tightly coupled systems create conflicts within modern distributed
architectures. For example, architects and DBAs must ensure that business data sur‐
vives the breaking apart of monolith systems and that the business can still derive
value from its data regardless of architecture undulations.
It has been said that data is the most important asset in a company. Businesses want to
extract value from the data that they have and are finding new ways to deploy data in
decision making. Every part of the enterprise is now data driven, from servicing exist‐
ing customers, to acquiring new customers, increasing customer retention, improv‐
ing products, predicting sales, and other trends. This reliance on data means that all
software architecture is in the service of data, ensuring the right data is available and
usable by all parts of the enterprise.
The authors built many distributed systems a few decades ago when they first became
popular, yet decision making in modern microservices seems more difficult, and we
wanted to figure out why. We eventually realized that, back in the early days of dis‐
tributed architecture, we mostly still persisted data in a single relational database.
However, in microservices and the philosophical adherence to a bounded context
from Domain-Driven Design, as a way of limiting the scope of implementation detail
coupling, data has moved to an architectural concern, along with transactionality.
Many of the hard parts of modern architecture derive from tensions between data
and architecture concerns, which we untangle in both Part I and Part II.
One important distinction that we cover in a variety of chapters is the separation
between operational versus analytical data:
Operational data
Data used for the operation of the business, including sales, transactional data,
inventory, and so on. This data is what the company runs on—if something inter‐
rupts this data, the organization cannot function for very long. This type of data
is defined as Online Transactional Processing (OLTP), which typically involves
inserting, updating, and deleting data in a database.
4
|
Chapter 1: What Happens When There Are No “Best Practices”?
Analytical data
Data used by data scientists and other business analysts for predictions, trending,
and other business intelligence. This data is typically not transactional and often
not relational—it may be in a graph database or snapshots in a different format
than its original transactional form. This data isn’t critical for the day-to-day
operation but rather for the long-term strategic direction and decisions.
We cover the impact of both operational and analytical data throughout the book.
Architectural Decision Records
One of the most effective ways of documenting architecture decisions is through
Architectural Decision Records (ADRs). ADRs were first evangelized by Michael
Nygard in a blog post and later marked as “adopt” in the Thoughtworks Technology
Radar. An ADR consists of a short text file (usually one to two pages long) describing
a specific architecture decision. While ADRs can be written using plain text, they are
usually written in some sort of text document format like AsciiDoc or Markdown.
Alternatively, an ADR can also be written using a wiki page template. We devoted an
entire chapter to ADRs in our previous book, Fundamentals of Software Architecture
(O’Reilly).
We will be leveraging ADRs as a way of documenting various architecture decisions
made throughout the book. For each architecture decision, we will be using the fol‐
lowing ADR format with the assumption that each ADR is approved:
ADR: A short noun phrase containing the architecture decision
Context
In this section of the ADR we will add a short one- or two-sentence description of the
problem, and list the alternative solutions.
Decision
In this section we will state the architecture decision and provide a detailed justifica‐
tion of the decision.
Consequences
In this section of the ADR we will describe any consequences after the decision is
applied, and also discuss the trade-offs that were considered.
A list of all the Architectural Decision Records created in this book can be found in
Appendix B.
Documenting a decision is important for an architect, but governing the proper use
of the decision is a separate topic. Fortunately, modern engineering practices allow
automating many common governance concerns by using architecture fitness
functions.
Architectural Decision Records
|
5
Architecture Fitness Functions
Once an architect has identified the relationship between components and codified
that into a design, how can they make sure that the implementers will adhere to that
design? More broadly, how can architects ensure that the design principles they
define become reality if they aren’t the ones to implement them?
These questions fall under the heading of architecture governance, which applies to
any organized oversight of one or more aspects of software development. As this
book primarily covers architecture structure, we cover how to automate design and
quality principles via fitness functions in many places.
Software development has slowly evolved over time to adapt unique engineering
practices. In the early days of software development, a manufacturing metaphor was
commonly applied to software practices, both in the large (like the Waterfall develop‐
ment process) and small (integration practices on projects). In the early 1990s, a
rethinking of software development engineering practices, lead by Kent Beck and the
other engineers on the C3 project, called eXtreme Programming (XP), illustrated the
importance of incremental feedback and automation as key enablers of software
development productivity. In the early 2000s, the same lessons were applied to the
intersection of software development and operations, spawning the new role of
DevOps and automating many formerly manual operational chores. Just as before,
automation allows teams to go faster because they don’t have to worry about things
breaking without good feedback. Thus, automation and feedback have become central
tenets for effective software development.
Consider the environments and situations that lead to breakthroughs in automation.
In the era before continuous integration, most software projects included a lengthy
integration phase. Each developer was expected to work in some level of isolation
from others, then integrate all the code at the end into an integration phase. Vestiges
of this practice still linger in version control tools that force branching and prevent
continuous integration. Not surprisingly, a strong correlation existed between project
size and the pain of the integration phase. By pioneering continuous integration, the
XP team illustrated the value of rapid, continuous feedback.
The DevOps revolution followed a similar course. As Linux and other open source
software became “good enough” for enterprises, combined with the advent of tools
that allowed programmatic definition of (eventually) virtual machines, operations
personnel realized they could automate machine definitions and many other repeti‐
tive tasks.
In both cases, advances in technology and insights led to automating a recurring job
that was handled by an expensive role—which describes the current state of architec‐
ture governance in most organizations. For example, if an architect chooses a particu‐
lar architecture style or communication medium, how can they make sure that a
6
|
Chapter 1: What Happens When There Are No “Best Practices”?
developer implements it correctly? When done manually, architects perform code
reviews or perhaps hold architecture review boards to assess the state of governance.
However, just as in manually configuring computers in operations, important details
can easily fall through superficial reviews.
Using Fitness Functions
In the 2017 book Building Evolutionary Architectures (O’Reilly), the authors (Neal
Ford, Rebecca Parsons, and Patrick Kua) defined the concept of an architectural fit‐
ness function: any mechanism that performs an objective integrity assessment of some
architecture characteristic or combination of architecture characteristics. Here is a
point-by-point breakdown of that definition:
Any mechanism
Architects can use a wide variety of tools to implement fitness functions; we will
show numerous examples throughout the book. For example, dedicated testing
libraries exist to test architecture structure, architects can use monitors to test
operational architecture characteristics such as performance or scalability, and
chaos engineering frameworks test reliability and resiliency.
Objective integrity assessment
One key enabler for automated governance lies with objective definitions for
architecture characteristics. For example, an architect can’t specify that they want
a “high performance” website; they must provide an object value that can be
measured by a test, monitor, or other fitness function.
Architects must watch out for composite architecture characteristics—ones that
aren’t objectively measurable but are really composites of other measurable
things. For example, “agility” isn’t measurable, but if an architect starts pulling
the broad term agility apart, the goal is for teams to be able to respond quickly
and confidently to change, either in ecosystem or domain. Thus, an architect can
find measurable characteristics that contribute to agility: deployability, testability,
cycle time, and so on. Often, the lack of ability to measure an architecture char‐
acteristic indicates too vague a definition. If architects strive toward measurable
properties, it allows them to automate fitness function application.
Some architecture characteristic or combination of architecture characteristics
This characteristic describes the two scopes for fitness functions:
Atomic
These fitness functions handle a single architecture characteristic in isola‐
tion. For example, a fitness function that checks for component cycles within
a codebase is atomic in scope.
Architecture Fitness Functions
|
7
Holistic
Holistic fitness functions validate a combination of architecture characteris‐
tics. A complicating feature of architecture characteristics is the synergy they
sometimes exhibit with other architecture characteristics. For example, if an
architect wants to improve security, a good chance exists that it will affect
performance. Similarly, scalability and elasticity are sometimes at odds—
supporting a large number of concurrent users can make handling sudden
bursts more difficult. Holistic fitness functions exercise a combination of
interlocking architecture characteristics to ensure that the combined effect
won’t negatively affect the architecture.
An architect implements fitness functions to build protections around unexpected
change in architecture characteristics. In the Agile software development world,
developers implement unit, functional, and user acceptance tests to validate different
dimensions of the domain design. However, until now, no similar mechanism existed
to validate the architecture characteristics part of the design. In fact, the separation
between fitness functions and unit tests provides a good scoping guideline for archi‐
tects. Fitness functions validate architecture characteristics, not domain criteria; unit
tests are the opposite. Thus, an architect can decide whether a fitness function or unit
test is needed by asking the question: “Is any domain knowledge required to execute
this test?” If the answer is “yes,” then a unit/function/user acceptance test is appropri‐
ate; if “no,” then a fitness function is needed.
For example, when architects talk about elasticity, it’s the ability of the application to
withstand a sudden burst of users. Notice that the architect doesn’t need to know any
details about the domain—this could be an ecommerce site, an online game, or some‐
thing else. Thus, elasticity is an architectural concern and within the scope of a fitness
function. If on the other hand the architect wanted to validate the proper parts of a
mailing address, that is covered via a traditional test. Of course, this separation isn’t
purely binary—some fitness functions will touch on the domain and vice versa, but
the differing goals provide a good way to mentally separate them.
Here are a couple of examples to make the concept less abstract.
One common architect goal is to maintain good internal structural integrity in the
codebase. However, malevolent forces work against the architect’s good intentions on
many platforms. For example, when coding in any popular Java or .NET development
environment, as soon as a developer references a class not already imported, the IDE
helpfully presents a dialog asking the developer if they would like to auto-import the
reference. This occurs so often that most programmers develop the habit of swatting
the auto-import dialog away like a reflex action.
However, arbitrarily importing classes or components among one another spells dis‐
aster for modularity. For example, Figure 1-1 illustrates a particularly damaging antipattern that architects aspire to avoid.
8
|
Chapter 1: What Happens When There Are No “Best Practices”?
Figure 1-1. Cyclic dependencies between components
In this anti-pattern, each component references something in the others. Having a
network of components such as this damages modularity because a developer cannot
reuse a single component without also bringing the others along. And, of course, if
the other components are coupled to other components, the architecture tends more
and more toward the Big Ball of Mud anti-pattern. How can architects govern this
behavior without constantly looking over the shoulders of trigger-happy developers?
Code reviews help but happen too late in the development cycle to be effective. If an
architect allows a development team to rampantly import across the codebase for a
week until the code review, serious damage has already occurred in the codebase.
The solution to this problem is to write a fitness function to avoid component cycles,
as shown in Example 1-1.
Example 1-1. Fitness function to detect component cycles
public class CycleTest {
private JDepend jdepend;
@BeforeEach
void init() {
jdepend = new JDepend();
jdepend.addDirectory("/path/to/project/persistence/classes");
jdepend.addDirectory("/path/to/project/web/classes");
jdepend.addDirectory("/path/to/project/thirdpartyjars");
}
@Test
void testAllPackages() {
Collection packages = jdepend.analyze();
assertEquals("Cycles exist", false, jdepend.containsCycles());
}
}
In the code, an architect uses the metrics tool JDepend to check the dependencies
between packages. The tool understands the structure of Java packages and fails the
test if any cycles exist. An architect can wire this test into the continuous build on a
project and stop worrying about the accidental introduction of cycles by triggerhappy developers. This is a great example of a fitness function guarding the impor‐
Architecture Fitness Functions
|
9
tant rather than urgent practices of software development: it’s an important concern
for architects, yet has little impact on day-to-day coding.
Example 1-1 shows a very low-level, code-centric fitness function. Many popular
code hygiene tools (such as SonarQube) implement many common fitness functions
in a turnkey manner. However, architects may also want to validate the macro struc‐
ture of the architecture as well as the micro. When designing a layered architecture
such as the one in Figure 1-2, the architect defines the layers to ensure separation of
concerns.
Figure 1-2. Traditional layered architecture
However, how can the architect ensure that developers will respect these layers? Some
developers may not understand the importance of the patterns, while others may
adopt a “better to ask forgiveness than permission” attitude because of some overrid‐
ing local concern, such as performance. But allowing implementers to erode the rea‐
sons for the architecture hurts the long-term health of the architecture.
ArchUnit allows architects to address this problem via a fitness function, shown in
Example 1-2.
Example 1-2. ArchUnit fitness function to govern layers
layeredArchitecture()
.layer("Controller").definedBy("..controller..")
.layer("Service").definedBy("..service..")
.layer("Persistence").definedBy("..persistence..")
.whereLayer("Controller").mayNotBeAccessedByAnyLayer()
.whereLayer("Service").mayOnlyBeAccessedByLayers("Controller")
.whereLayer("Persistence").mayOnlyBeAccessedByLayers("Service")
10
|
Chapter 1: What Happens When There Are No “Best Practices”?
In Example 1-2, the architect defines the desirable relationship between layers and
writes a verification fitness function to govern it. This allows an architect to establish
architecture principles outside the diagrams and other informational artifacts, and
verify them on an ongoing basis.
A similar tool in the .NET space, NetArchTest, allows similar tests for that platform.
A layer verification in C# appears in Example 1-3.
Example 1-3. NetArchTest for layer dependencies
// Classes in the presentation should not directly reference repositories
var result = Types.InCurrentDomain()
.That()
.ResideInNamespace("NetArchTest.SampleLibrary.Presentation")
.ShouldNot()
.HaveDependencyOn("NetArchTest.SampleLibrary.Data")
.GetResult()
.IsSuccessful;
Tools continue to appear in this space with increasing degrees of sophistication. We
will continue to highlight many of these techniques as we illustrate fitness functions
alongside many of our solutions.
Finding an objective outcome for a fitness function is critical. However, objective
doesn’t imply static. Some fitness functions will have noncontextual return values,
such as true/false or a numeric value such as a performance threshold. However,
other fitness functions (deemed dynamic) return a value based on some context. For
example, when measuring scalability, architects measure the number of concurrent
users and also generally measure the performance for each user. Often, architects
design systems so that as the number of users goes up, performance per user declines
slightly—but doesn’t fall off a cliff. Thus, for these systems, architects design perfor‐
mance fitness functions that take into account the number of concurrent users. As
long as the measure of an architecture characteristic is objective, architects can test it.
While most fitness functions should be automated and run continually, some will
necessarily be manual. A manual fitness function requires a person to handle the vali‐
dation. For example, for systems with sensitive legal information, a lawyer may need
to review changes to critical parts to ensure legality, which cannot be automated.
Most deployment pipelines support manual stages, allowing teams to accommodate
manual fitness functions. Ideally, these are run as often as reasonably possible—a vali‐
dation that doesn’t run can’t validate anything. Teams execute fitness functions either
on demand (rarely) or as part of a continuous integration work stream (most
common). To fully achieve the benefit of validations such as fitness functions, they
should be run continually.
Architecture Fitness Functions
|
11
Continuity is important, as illustrated in this example of enterprise-level governance
using fitness functions. Consider the following scenario: what does a company do
when a zero-day exploit is discovered in one of the development frameworks or libra‐
ries the enterprise uses? If it’s like most companies, security experts scour projects to
find the offending version of the framework and make sure it’s updated, but that pro‐
cess is rarely automated, relying on many manual steps. This isn’t an abstract ques‐
tion; this exact scenario affected a major financial institution described in The
Equifax Data Breach. Like the architecture governance described previously, manual
processes are error prone and allow details to escape.
The Equifax Data Breach
On September 7, 2017, Equifax, a major credit scoring agency in the US, announced
that a data breach had occurred. Ultimately, the problem was traced to a hacking
exploit of the popular Struts web framework in the Java ecosystem (Apache Struts
vCVE-2017-5638). The foundation issued a statement announcing the vulnerability
and released a patch on March 7, 2017. The Department of Homeland Security con‐
tacted Equifax and similar companies the next day, warning them of this problem,
and they ran scans on March 15, 2017, which didn’t reveal all of the affected systems.
Thus, the critical patch wasn’t applied to many older systems until July 29, 2017, when
Equifax’s security experts identified the hacking behavior that lead to the data breach.
Imagine an alternative world in which every project runs a deployment pipeline, and
the security team has a “slot” in each team’s deployment pipeline where they can
deploy fitness functions. Most of the time, these will be mundane checks for safe‐
guards like preventing developers from storing passwords in databases and similar
regular governance chores. However, when a zero-day exploit appears, having the
same mechanism in place everywhere allows the security team to insert a test in every
project that checks for a certain framework and version number; if it finds the dan‐
gerous version, it fails the build and notifies the security team. Teams configure
deployment pipelines to awaken for any change to the ecosystem: code, database
schema, deployment configuration, and fitness functions. This allows enterprises to
universally automate important governance tasks.
Fitness functions provide many benefits for architects, not the least of which is the
chance to do some coding again! One of the universal complaints among architects is
that they don’t get to code much anymore—but fitness functions are often code! By
building an executable specification of the architecture, which anyone can validate
anytime by running the project’s build, architects must understand the system and its
ongoing evolution well, which overlaps with the core goal of keeping up with the
code of the project as it grows.
12
|
Chapter 1: What Happens When There Are No “Best Practices”?
However powerful fitness functions are, architects should avoid overusing them.
Architects should not form a cabal and retreat to an ivory tower to build an impossi‐
bly complex, interlocking set of fitness functions that merely frustrate developers and
teams. Instead, it’s a way for architects to build an executable checklist of important
but not urgent principles on software projects. Many projects drown in urgency,
allowing some important principles to slip by the side. This is the frequent cause of
technical debt: “We know this is bad, but we’ll come back to fix it later”—and later
never comes. By codifying rules about code quality, structure, and other safeguards
against decay into fitness functions that run continually, architects build a quality
checklist that developers can’t skip.
A few years ago, the excellent book The Checklist Manifesto by Atul Gawande (Pica‐
dor) highlighted the use of checklists by professionals like surgeons, airline pilots,
and those other fields who commonly use (sometimes by force of law) checklists as
part of their job. It isn’t because they don’t know their job or are particularly forgetful;
when professionals perform the same task over and over, it becomes easy to fool
themselves when it’s accidentally skipped, and checklists prevent that. Fitness func‐
tions represent a checklist of important principles defined by architects and run as
part of the build to make sure developers don’t accidentally (or purposefully, because
of external forces like schedule pressure) skip them.
We utilize fitness functions throughout the book when an opportunity arises to illus‐
trate governing an architectural solution as well as the initial design.
Architecture Versus Design: Keeping Definitions Simple
A constant area of struggle for architects is keeping architecture and design as separate
but related activities. While we don’t want to wade into the never-ending argument
about this distinction, we strive in this book to stay firmly on the architecture side of
that spectrum for several reasons.
First, architects must understand underlying architecture principles to make effective
decisions. For example, the decision between synchronous versus asynchronous com‐
munication has a number of trade-offs before architects layer in implementation
details. In the book Fundamentals of Software Architecture, the authors coined the
second law of software architecture: why is more important than how. While ulti‐
mately architects must understand how to implement solutions, they must first
understand why one choice has better trade-offs than another.
Second, by focusing on architecture concepts, we can avoid the numerous implemen‐
tations of those concepts. Architects can implement asynchronous communication in
a variety of ways; we focus on why an architect would choose asynchronous commu‐
nication and leave the implementation details to another place.
Architecture Versus Design: Keeping Definitions Simple
|
13
Third, if we start down the path of implementing all the varieties of options we show,
this would be the longest book ever written. Focus on architecture principles allows
us to keep things as generic as they can be.
To keep subjects as grounded in architecture as possible, we use the simplest defini‐
tions possible for key concepts. For example, coupling in architecture can fill entire
books (and it has). To that end, we use the following simple, verging on simplistic,
definitions:
Service
In colloquial terms, a service is a cohesive collection of functionality deployed as
an independent executable. Most of the concepts we discuss with regard to serv‐
ices apply broadly to distributed architectures, and specifically microservices
architectures.
In the terms we define in Chapter 2, a service is part of an architecture quantum,
which includes further definitions of both static and dynamic coupling between
services and other quanta.
Coupling
Two artifacts (including services) are coupled if a change in one might require a
change in the other to maintain proper functionality.
Component
An architectural building block of the application that does some sort of business
or infrastructure function, usually manifested through a package structure (Java),
namespace (C#), or a physical grouping of source code files within some sort of
directory structure. For example, the component Order History might be imple‐
mented through a set of class files located in the namespace app.busi
ness.order.history.
Synchronous communication
Two artifacts communicate synchronously if the caller must wait for the response
before proceeding.
Asynchronous communication
Two artifacts communicate asynchronously if the caller does not wait for the
response before proceeding. Optionally, the caller can be notified by the receiver
through a separate channel when the request has completed.
Orchestrated coordination
A workflow is orchestrated if it includes a service whose primary responsibility is
to coordinate the workflow.
14
|
Chapter 1: What Happens When There Are No “Best Practices”?
Choreographed coordination
A workflow is choreographed when it lacks an orchestrator; rather, the services
in the workflow share the coordination responsibilities of the workflow.
Atomicty
A workflow is atomic if all parts of the workflow maintain a consistent state at all
times; the opposite is represented by the spectrum of eventual consistency, cov‐
ered in Chapter 6.
Contract
We use the term contract broadly to define the interface between two software
parts, which may encompass method or function calls, integration architecture
remote calls, dependencies, and so on. Anywhere two pieces of software join, a
contract is involved.
Software architecture is by its nature abstract: we cannot know what unique combi‐
nation of platforms, technologies, commercial software, and the other dizzying array
of possibilities our readers might have, except that no two are exactly alike. We cover
many abstract ideas, but must ground them with some implementation details to
make them concrete. To that end, we need a problem to illustrate architecture con‐
cepts against—which leads us to the Sysops Squad.
Introducing the Sysops Squad Saga
saga
A long story of heroic achievement.
—Oxford English Dictionary
We discuss a number of sagas in this book, both literal and figurative. Architects have
co-opted the term saga to describe transactional behavior in distributed architectures
(which we cover in detail in Chapter 12). However, discussions about architecture
tend to become abstract, especially when considering abstract problems such as the
hard parts of architecture. To help solve this problem and provide some real-world
context for the solutions we discuss, we kick off a literal saga about the Sysops Squad.
We use the Sysops Squad saga within each chapter to illustrate the techniques and
trade-offs described in this book. While many books on software architecture cover
new development efforts, many real-world problems exist within existing systems.
Therefore, our story starts with the existing Sysops Squad architecture highlighted
here.
Penultimate Electronics is a large electronics giant that has numerous retail stores
throughout the country. When customers buy computers, TVs, stereos, and other
electronic equipment, they can choose to purchase a support plan. When problems
Introducing the Sysops Squad Saga
|
15
occur, customer-facing technology experts (the Sysops Squad) come to the customer’s
residence (or work office) to fix problems with the electronic device.
The four main users of the Sysops Squad ticketing application are as follows:
Administrator
The administrator maintains the internal users of the system, including the list of
experts and their corresponding skill set, location, and availability. The adminis‐
trator also manages all of the billing processing for customers using the system,
and maintains static reference data (such as supported products, name-value
pairs in the system, and so on).
Customer
The customer registers for the Sysops Squad service and maintains their cus‐
tomer profile, support contracts, and billing information. Customers enter prob‐
lem tickets into the system, and also fill out surveys after the work has been
completed.
Sysops Squad expert
Experts are assigned problem tickets and fix problems based on the ticket. They
also interact with the knowledge base to search for solutions to customer prob‐
lems and enter notes about repairs.
Manager
The manager keeps track of problem ticket operations and receives operational
and analytical reports about the overall Sysops Squad problem ticket system.
Nonticketing Workflow
The nonticketing workflows include those actions that administrators, managers, and
customers perform that do not relate to a problem ticket. These workflows are out‐
lined as follows:
1. Sysops Squad experts are added and maintained in the system through an
administrator, who enters in their locale, availability, and skills.
2. Customers register with the Sysops Squad system and have multiple support
plans based on the products they purchased.
3. Customers are automatically billed monthly based on credit card information
contained in their profile. Customers can view billing history and statements
through the system.
4. Managers request and receive various operational and analytical reports, includ‐
ing financial reports, expert performance reports, and ticketing reports.
16
|
Chapter 1: What Happens When There Are No “Best Practices”?
Ticketing Workflow
The ticketing workflow starts when a customer enters a problem ticket into the sys‐
tem, and ends when the customer completes the survey after the repair is done. This
workflow is outlined as follows:
1. Customers who have purchased the support plan enter a problem ticket by using
the Sysops Squad website.
2. Once a problem ticket is entered in the system, the system then determines
which Sysops Squad expert would be the best fit for the job based on skills, cur‐
rent location, service area, and availability.
3. Once assigned, the problem ticket is uploaded to a dedicated custom mobile app
on the Sysops Squad expert’s mobile device. The expert is also notified via a text
message that they have a new problem ticket.
4. The customer is notified through an SMS text message or email (based on their
profile preference) that the expert is on their way.
5. The expert uses the custom mobile application on their phone to retrieve the
ticket information and location. The Sysops Squad expert can also access a
knowledge base through the mobile app to find out what has been done in the
past to fix the problem.
6. Once the expert fixes the problem, they mark the ticket as “complete.” The sysops
squad expert can then add information about the problem and repair the knowl‐
edge base.
7. After the system receives notification that the ticket is complete, it sends an email
to the customer with a link to a survey, which the customer then fills out.
8. The system receives the completed survey from the customer and records the
survey information.
A Bad Scenario
Things have not been good with the Sysops Squad problem ticket application lately.
The current trouble ticket system is a large monolithic application that was developed
many years ago. Customers are complaining that consultants are never showing up
because of lost tickets, and often the wrong consultant shows up to fix something
they know nothing about. Customers have also been complaining that the system is
not always available to enter new problem tickets.
Change is also difficult and risky in this large monolith. Whenever a change is made,
it usually takes too long and something else usually breaks. Because of reliability
issues, the Sysops Squad system frequently “freezes up,” or crashes, resulting in all
application functionality not being available anywhere from five minutes to two
hours while the problem is identified and the application restarted.
Introducing the Sysops Squad Saga
|
17
If something isn’t done soon, Penultimate Electronics will be forced to abandon the
very lucrative support contract business line and lay off all the Sysops Squad adminis‐
trators, experts, managers, and IT development staff—including the architects.
Sysops Squad Architectural Components
The monolithic system for the Sysops Squad application handles ticket management,
operational reporting, customer registration, and billing, as well as general adminis‐
trative functions such as user maintenance, login, and expert skills and profile main‐
tenance. Figure 1-3 and the corresponding Table 1-1 illustrate and describe the
components of the existing monolithic application (the ss. part of the namespace
specifies the Sysops Squad application context).
Figure 1-3. Components within the existing Sysops Squad application
18
|
Chapter 1: What Happens When There Are No “Best Practices”?
Table 1-1. Existing Sysops Squad components
Component
Namespace
Responsibility
Login
ss.login
Internal user and customer login and security logic
Billing payment
ss.billing.payment
Customer monthly billing and customer credit card info
Billing history
ss.billing.history
Payment history and prior billing statements
Customer notification
ss.customer.notifica
tion
Notify customer of billing, general info
Customer profile
ss.customer.profile
Maintain customer profile, customer registration
Expert profile
ss.expert.profile
Maintain expert profile (name, location, skills, etc.)
KB maint
ss.kb.maintenance
Maintain and view items in the knowledge base
KB search
ss.kb.search
Query engine for searching the knowledge base
Reporting
ss.reporting
All reporting (experts, tickets, financial)
Ticket
ss.ticket
Ticket creation, maintenance, completion, common code
Ticket assign
ss.ticket.assign
Find an expert and assign the ticket
Ticket notify
ss.ticket.notify
Notify customer that the expert is on their way
Ticket route
ss.ticket.route
Send the ticket to the expert’s mobile device app
Support contract
ss.supportcontract
Support contracts for customers, products in the plan
Survey
ss.survey
Maintain surveys, capture and record survey results
Survey notify
ss.survey.notify
Send survey email to customer
Survey templates
ss.survey.templates
Maintain various surveys based on type of service
User maintenance
ss.users
Maintain internal users and roles
These components will be used in subsequent chapters to illustrate various techni‐
ques and trade-offs when dealing with breaking applications into distributed
architectures.
Sysops Squad Data Model
The Sysops Squad application with its various components listed in Table 1-1 uses a
single schema in the database to host all its tables and related database code. The
database is used to persist customers, users, contracts, billing, payments, knowledge
base, and customer surveys; the tables are listed in Table 1-2, and the ER model is
illustrated in Figure 1-4.
Introducing the Sysops Squad Saga
|
19
Figure 1-4. Data model within the existing Sysops Squad application
Table 1-2. Existing Sysops Squad database tables
Table
Responsibility
Customer
Customer_Notification
Survey
Question
Survey_Question
Survey_Administered
Survey_Response
Billing
Contract
Payment_Method
Payment
Entities needing Sysops support
Notification preferences for customers
A survey for after-support customer satisfaction
Questions in a survey
A question is assigned to the survey
Survey question is assigned to customer
A customer’s response to the survey
Billing information for support contract
A contract between an entity and Sysops for support
Payment methods supported for making payment
Payments processed for billings
20
|
Chapter 1: What Happens When There Are No “Best Practices”?
Table
Responsibility
SysOps_User
Profile
Expert_Profile
Expertise
Location
Article
Tag
Keyword
Article_Tag
Article_Keyword
Ticket
Ticket_Type
Ticket_History
The various users in Sysops
Profile information for Sysops users
Profiles of experts
Various expertise within Sysops
Locations served by the expert
Articles for the knowledge base
Tags on articles
Keyword for an article
Tags associated to articles
Join table for keywords and articles
Support tickets raised by customers
Different types of tickets
The history of support tickets
The Sysops data model is a standard third normal form data model with only a few
stored procedures or triggers. However, a fair number of views exist that are mainly
used by the Reporting component. As the architecture team tries to break up the
application and move toward distributed architecture, it will have to work with the
database team to accomplish the tasks at the database level. This setup of database
tables and views will be used throughout the book to discuss various techniques and
trade-offs to accomplish the task of breaking apart the database.
Introducing the Sysops Squad Saga
|
21
PART I
Pulling Things Apart
As many of us discovered when we were children, a great way to understand how
something fits together is to first pull it apart. To understand complex subjects (such
as trade-offs in distributed architectures), an architect must figure out where to start
untangling.
In the book What Every Programmer Should Know About Object-Oriented Design
(Dorset House), Meilir Page-Jones made the astute observation that coupling in
architecture may be split into static and dynamic coupling. Static coupling refers to
the way architectural parts (classes, components, services, and so on) are wired
together: dependencies, coupling degree, connection points, and so on. An architect
can often measure static coupling at compile time as it represents the static depen‐
dencies within the architecture.
Dynamic coupling refers to how architecture parts call one another: what kind of
communication, what information is passed, strictness of contracts, and so on.
Our goal is to investigate how to do trade-off analysis in distributed architectures; to
do that, we must pull the moving pieces apart so that we can discuss them in isolation
to understand them fully before putting them back together.
Part I primarily deals with architectural structure, how things are statically coupled
together. In Chapter 2, we tackle the problem of defining the scope of static and
dynamic coupling in architectures, and present the entire picture that we must pull
apart to understand. Chapter 3 begins that process, defining modularity and separa‐
tion in architecture. Chapter 4 provides tools to evaluate and deconstruct codebases,
and Chapter 5 supplies patterns to assist the process.
Data and transactions have become increasingly important in architecture, driving
many trade-off decisions by architects and DBAs. Chapter 6 addresses the architec‐
tural impacts of data, including how to reconcile service and data boundaries. Finally,
Chapter 7 ties together architecture coupling with data concerns to define integrators
and disintegrators—forces that encourage a larger or smaller service size and
boundary.
CHAPTER 2
Discerning Coupling in
Software Architecture
Wednesday, November 3, 13:00
Logan, the lead architect for Penultimate Electronics, interrupted a small group of architects in the cafeteria, discussing distributed architectures. “Austen, are you wearing a cast
again?”
“No, it’s just a splint,” replied Austen. “I sprained my wrist playing extreme disc golf over
the weekend—it’s almost healed.”
“What is…never mind. What is this impassioned conversation I barged in on?”
“Why wouldn’t someone always choose the saga pattern in microservices to wire together transactions?” asked Austen. “That way, architects can make the services as small as they want.”
“But don’t you have to use orchestration with sagas?” asked Addison. “What about times when we
need asynchronous communication? And, how complex will the transactions get? If we break things
down too much, can we really guarantee data fidelity?”
“You know,” said Austen, “if we use an enterprise service bus, we can get it to manage most of that
stuff for us.”
“I thought no one used ESBs anymore—shouldn’t we use Kafka for stuff like that?”
“They aren’t even the same thing!” said Austen.
Logan interrupted the increasingly heated conversation. “It is an apples-to-oranges comparison, but
none of these tools or approaches is a silver bullet. Distributed architectures like microservices are
difficult, especially if architects cannot untangle all the forces at play. What we need is an approach
or framework that helps us figure out the hard problems in our architecture.”
25
“Well,” said Addison, “whatever we do, it has to be as decoupled as possible—everything I’ve read
says that architects should embrace decoupling as much as possible.”
“If you follow that advice,” said Logan, “Everything will be so decoupled that nothing can communicate with anything else—it’s hard to build software that way! Like a lot of things, coupling isn’t inherently bad; architects just have to know how to apply it appropriately. In fact, I remember a famous
quote about that from a Greek philosopher….”
All things are poison, and nothing is without poison; the dosage alone makes it so a
thing is not a poison.
—Paracelsus
One of the most difficult tasks an architect will face is untangling the various forces
and trade-offs at play in distibuted architectures. People who provide advice con‐
stantly extol the benefits of “loosely coupled” systems, but how can architects design
systems where nothing connects to anything else? Architects design fine-grained
microservices to achieve decoupling, but then orchestration, transactionality, and
asynchronicity become huge problems. Generic advice says “decouple,” but provides
no guidelines for how to achieve that goal while still constructing useful systems.
Architects struggle with granularity and communication decisions because there are
no clear universal guides for making decisions—no best practices exist that can apply
to real-world complex systems. Until now, architects lacked the correct perspective
and terminology to allow a careful analysis that could determine the best (or least
worst) set of trade-offs on a case-by-case basis.
Why have architects struggled with decisions in distributed architectures? After all,
we’ve been building distributed systems since the last century, using many of the
same mechanisms (message queues, events, and so on). Why has the complexity ram‐
ped up so much with microservices?
The answer lies with the fundamental philosophy of microservices, inspired by the
idea of a bounded context. Building services that model bounded contexts required a
subtle but important change to the way architects designed distributed systems
because now transactionality is a first-class architectural concern. In many of the dis‐
tributed systems architects designed prior to microservices, event handlers typically
connected to a single relational database, allowing it to handle details such as integ‐
rity and transactions. Moving the database within the service boundary moves data
concerns into architecture concerns.
As we’ve said before, “Software architecture” is the stuff you can’t Google answers for. A
skill that modern architects must build is the ability to do trade-off analysis. While
several frameworks have existed for decades (such as Architecture Trade-off Analysis
Method, or ATAM), they lack focus on real problems architects face on a daily basis.
26
|
Chapter 2: Discerning Coupling in Software Architecture
This book focuses on how architects can perform trade-off analysis for any number
of scenarios unique to their situation. As in many things in architecture, the advice is
simple; the hard parts lie in the details, particularly how difficult parts become entan‐
gled, making it difficult to see and understand the individual parts, as illustrated in
Figure 2-1.
Figure 2-1. A braid entangles hair, making the individual strands hard to identify
When architects look at entangled problems, they struggle with performing trade-off
analysis because of the difficulties separating the concerns, so that they may consider
them independently. Thus, the first step in trade-off analysis is untangle the dimen‐
sions of the problem, analyzing what parts are coupled to one another and what
impact that coupling has on change. For this purpose, we use the simplest definition
of the word coupling:
Coupling
Two parts of a software system are coupled if a change in one might cause a
change in the other.
Often, software architecture creates multidimensional problems, where multiple
forces all interact in interdependent ways. To analyze trade-offs, an architect must
first determine what forces need to trade off with each other.
Thus, here’s our advice for modern trade-off analysis in software architecture:
1. Find what parts are entangled together.
2. Analyze how they are coupled to one another.
3. Assess trade-offs by determining the impact of change on interdependent
systems.
While the steps are simple, the hard parts lurk in the details. Thus, to illustrate this
framework in practice, we take one of the most difficult (and probably the closest to
generic) problems in distributed architectures, which is related to microservices:
Discerning Coupling in Software Architecture
|
27
How do architects determine the size and communication styles for microservices?
Determining the proper size for microservices seems a pervasive problem—toosmall services create transactional and orchestration issues, and too-large serv‐
ices create scale and distribution issues.
To that end, the remainder of this book untangles the many aspects to consider when
answering the preceding question. We provide new terminology to differentiate simi‐
lar but distinct patterns and show practical examples of applying these and other
patterns.
However, the overarching goal of this book is to provide you with example-driven
techniques to learn how to construct your own trade-off analysis for the unique prob‐
lems within your realm. We start with our first great untangling of forces in dis‐
tributed architectures: defining architecture quantum along with the two types of
coupling, static and dynamic.
Architecture (Quantum | Quanta)
The term quantum is, of course, used heavily in the field of physics known as quan‐
tum mechanics. However, the authors chose the word for the same reasons physicists
did. Quantum originated from the Latin word quantus, meaning “how great” or “how
many.” Before physics co-opted it, the legal profession used it to represent the
“required or allowed amount” (for example, in damages paid). The term also appears
in the mathematics field of topology, concerning the properties of families of shapes.
Because of its Latin roots, the singular is quantum, and the plural is quanta, similar to
the datum/data symmetry.
An architecture quantum measures several aspects of both topology and behavior in
software architecture related to how parts connect and communicate with one
another:
Architecture quantum
An architecture quantum is an independently deployable artifact with high func‐
tional cohesion, high static coupling, and synchronous dynamic coupling. A
common example of an architecture quantum is a well-formed microservice
within a workflow.
Static coupling
Represents how static dependencies resolve within the architecture via contracts.
These dependencies include operating system, frameworks, and/or libraries
delivered via transitive dependency management, and any other operational
requirement to allow the quantum to operate.
28
|
Chapter 2: Discerning Coupling in Software Architecture
Dynamic coupling
Represents how quanta communicate at runtime, either synchronously or asyn‐
chronously. Thus, fitness functions for these characteristics must be continuous,
typically utilizing monitors.
Even though both static and dynamic coupling seem similar, architects must distin‐
guish two important differences. An easy way to think about the difference is that
static coupling describes how services are wired together, whereas dynamic coupling
describes how services call one another at runtime. For example, in a microservices
architecture, a service must contain dependent components such as a database, repre‐
senting static coupling—the service isn’t operational without the necessary data. That
service may call other services during the course of a workflow, which represents
dynamic coupling. Neither service requires the other to be present to function, except
for this runtime workflow. Thus, static coupling analyzes operational dependencies,
and dynamic coupling analyzes communication dependencies.
These definitions include important characteristics; let’s cover each in detail as they
inform most of the examples in the book.
Independently Deployable
Independently deployable implies several aspects of an architecture quantum—each
quantum represents a separate deployable unit within a particular architecture. Thus,
a monolithic architecture—one that is deployed as a single unit—is by definition a
single architecture quantum. Within a distributed architecture such as microservices,
developers tend toward the ability to deploy services independently, often in a highly
automated way. Thus, from an independently deployable standpoint, a service within
a microservices architecture represents an architecture quantum (contingent on cou‐
pling—as discussed next).
Making each architecture quantum represent a deployable asset within the architec‐
ture serves several useful purposes. First, the boundary represented by an architec‐
ture quantum serves as a useful common language among architects, developers, and
operations. Each understands the common scope under question: architects under‐
stand the coupling characteristics, developers understand the scope of behavior, and
the operations team understands the deployable characteristics.
Second, the architecture quantum represents one of the forces (static coupling) archi‐
tects must consider when striving for proper granularity of services within a dis‐
tributed architecture. Often, in microservices architectures, developers face the
difficult question of what service granularity offers the optimum set of trade-offs.
Some of those trade-offs revolve around deployability: what release cadence does this
service require, what other services might be affected, what engineering practices are
involved, and so on. Architects benefit from a firm understanding of exactly where
Architecture (Quantum | Quanta)
|
29
deployment boundaries lie in distributed architectures. We discuss service granularity
and its attendant trade-offs in Chapter 7.
Third, independent deployability forces the architecture quantum to include common
coupling points such as databases. Most discussions about architecture conveniently
ignore issues such as databases and user interfaces, but real-world systems must com‐
monly deal with those problems. Thus, any system that uses a shared database fails
the architecture quantum criteria for independent deployment unless the database
deployment is in lockstep with the application. Many distributed systems that would
otherwise qualify for multiple quanta fail the independently deployable part if they
share a common database that has its own deployment cadence. Thus, merely consid‐
ering the deployment boundaries doesn’t solely provide a useful measure. Architects
should also consider the second criteria for an architecture quantum, high functional
cohesion, to limit the architecture quantum to a useful scope.
High Functional Cohesion
High functional cohesion refers structurally to the proximity of related elements:
classes, components, services, and so on. Throughout history, computer scientists
defined a variety of cohesion types, scoped in this case to the generic module, which
may be represented as classes or components, depending on platform. From a domain
standpoint, the technical definition of high functional cohesion overlaps with the goals
of the bounded context in domain-driven design: behavior and data that implements a
particular domain workflow.
From a purely independent deployability standpoint, a giant monolithic architecture
qualifies as an architecture quantum. However, it almost certainly isn’t highly func‐
tionally cohesive, but rather includes the functionality of the entire system. The larger
the monolith, the less likely it is singularly functionally cohesive.
Ideally, in a microservices architecture, each service models a single domain or work‐
flow, and therefore exhibits high functional cohesion. Cohesion in this context isn’t
about how services interact to perform work, but rather how independent and cou‐
pled one service is to another service.
High Static Coupling
High static coupling implies that the elements inside the architecture quantum are
tightly wired together, which is really an aspect of contracts. Architects recognize
things like REST or SOAP as contract formats, but method signatures and opera‐
tional dependencies (via coupling points such as IP addresses or URLs) also represent
contracts. Thus, contracts are an architecture hard part; we cover coupling issues
involving all types of contracts, including how to choose appropriate ones, in
Chapter 13.
30
|
Chapter 2: Discerning Coupling in Software Architecture
An architecture quantum is, in part, a measure of static coupling, and the measure is
quite simple for most architecture topologies. For example, the following diagrams
show the architecture styles featured in Fundamentals of Software Architecture, with
the architecture quantum static coupling illustrated.
Any of the monolithic architecture styles will necessarily have a quantum of one, as
illustrated in Figure 2-2.
Figure 2-2. Monolithic architectures always have a quantum of one
As you can see, any architecture that deploys as a single unit and utilizes a single
database will always have a single quantum. The architecture quantum measure of
static coupling includes the database, and a system that relies on a single database
cannot have more than a single quantum. Thus, the static coupling measure of an
architecture quantum helps identify coupling points in architecture, not just within
Architecture (Quantum | Quanta)
|
31
the software components under development. Most monolithic architectures contain
a single coupling point (typically, a database) that makes its quantum measure one.
Distributed architectures often feature decoupling at the component level; consider
the next set of architecture styles, starting with the service-based architecture shown
in Figure 2-3.
Figure 2-3. Architecture quantum for a service-based architecture
While this individual services model shows the isolation common in microservices,
the architecture still utilizes a single relational database, rendering its architecture
quantum score to one.
Service-Based Architecture
When we refer to service-based architecture, we don’t mean a generic architecture
based on services, but rather a specific hybrid architecture style that follows a dis‐
tributed macro-layered structure consisting of a separately deployed user interface,
separately deployed remote coarse-grained services, and a monolithic database. This
architecture addresses one of the complexities of microservices—separation at the
database level. Services in a service-based architecture follow the same principles as
microservices (based on domain-driven design’s bounded context) but rely on a single
relational database because the architects didn’t see value in separation (or saw too
many negative trade-offs).
Service-based architectures are common targets when restructuring monolithic archi‐
tectures, allowing for decomposition without disrupting existing database schemas
and integration points. We cover decomposition patterns in Chapter 5.
32
|
Chapter 2: Discerning Coupling in Software Architecture
So far, the static coupling measurement of architecture quantum has evaluated all the
topologies to one. However, distributed architectures create the possibility of multiple
quanta but don’t necessarily guarantee it. For example, the mediator style of eventdriven architecture will always be evaluated to an single architecture quantum, as
illustrated in Figure 2-4.
Even though this style represents a distributed architecture, two coupling points push
it toward a single architecture quantum: the database, as common with the previous
monolithic architectures, but also the Request Orchestrator itself—any holistic
coupling point necessary for the architecture to function forms an architecture quan‐
tum around it.
Figure 2-4. A mediated EDA has a single architecture quantum
Broker event-driven architectures (without a central mediator) are less coupled, but
that doesn’t guarantee complete decoupling. Consider the event-driven architecture
illustrated in Figure 2-5.
This broker-style event driven architecture (without a central mediator) is neverthe‐
less a single architecture quantum because all the services utilize a single relational
database, which acts as a common coupling point. The question answered by the
static analysis for an architecture quantum is, “Is this dependent of the architecture
necessary to bootstrap this service?” Even in the case of an event-driven architecture
where some of the services don’t access the database, if they rely on services that do
access the database, then they become part of the static coupling of the architecture
quantum.
Architecture (Quantum | Quanta)
|
33
Figure 2-5. Even a distributed architecture such as broker-style event-driven architecture
can be a single quantum
However, what about situations in distributed architectures where common coupling
points don’t exist? Consider the event-driven architecture illustrated in Figure 2-6.
The architects designed this event-driven system with two data stores, and no static
dependencies between the sets of services. Note that either architecture quantum can
run in a production-like ecosystem. It may not be able to participate in all workflows
required by the system, but it runs successfully and operates—sends requests and
receives them within the architecture.
The static coupling measure of an architecture quantum assesses the coupling depen‐
dencies between architectural and operational components. Thus, the operating sys‐
tem, data store, message broker, container orchestration, and all other operational
dependencies form the static coupling points of an architecture quantum, using the
strictest possible contracts, operational dependencies (more about the role of con‐
tracts in architecture quanta in Chapter 13).
34
|
Chapter 2: Discerning Coupling in Software Architecture
Figure 2-6. An event-driven architecture with multiple quanta
The microservices architecture style features highly decoupled services, including
data dependencies. Architects in these architectures favor high degrees of decoupling
and take care not to create coupling points between services, allowing each individual
service to each form its own quanta, as shown in Figure 2-7.
Figure 2-7. Microservices may form their own quanta
Architecture (Quantum | Quanta)
|
35
Each service (acting as a bounded context) may have its own set of architecture char‐
acteristics—one service might have higher levels of scalability or security than
another. This granular level of architecture characteristics scoping represents one of
the advantages of the microservices architecture style. High degrees of decoupling
allow teams working on a service to move as quickly as possible, without worrying
about breaking other dependencies.
However, if the system is tightly coupled to a user interface, the architecture forms a
single architecture quantum, as illustrated in Figure 2-8.
Figure 2-8. A tightly coupled user interface can reduce a microservices architecture
quantum to one
User interfaces create coupling points between the front and back end, and most user
interfaces won’t operate if portions of the backend aren’t available.
Additionally, it will be difficult for an architect to design different levels of opera‐
tional architecture characteristics (performance, scale, elasticity, reliability, and so on)
for each service if they all must cooperate together in a single user interface (particu‐
larly in the case of synchronous calls, covered in “Dynamic Quantum Coupling” on
page 38).
Architects design user interfaces utilizing asynchronicity that doesn’t create coupling
between front and back. A trend on many microservices projects is to use a micro
frontend framework for user interface elements in a microservices architecture. In
such an architecture, the user interface elements that interact on behalf of the services
are emitted from the services themselves. The user interface surface acts as a canvas
36
|
Chapter 2: Discerning Coupling in Software Architecture
where the user interface elements can appear, and also facilitates loosely coupled
communication between components, typically using events. Such an architecture is
illustrated in Figure 2-9.
Figure 2-9. In a micro-frontend architecture, each service + user interface component
forms an architecture quantum
In this example, the four tinted services along with their corresponding microfrontends form architecture quanta: each of these services may have different archi‐
tecture characteristics.
Any coupling point in an architecture can create static coupling points from a quan‐
tum standpoint. Consider the impact of a shared database between two systems, as
illustrated in Figure 2-10.
The static coupling of a system provides valuable insight, even in complex systems
involving integration architecture. Increasingly, a common architect technique for
understanding legacy architecture involves creating a static quantum diagram of how
things are “wired” together, which helps determine what systems will be impacted by
change and offers a way of understanding (and potentially decoupling) the
architecture.
Static coupling is only one-half of the forces at play in distributed architectures. The
other is dynamic coupling.
Architecture (Quantum | Quanta)
|
37
Figure 2-10. A shared database forms a coupling point between two systems, creating a
single quantum
Dynamic Quantum Coupling
The last portion of the architecture quantum definition concerns synchronous cou‐
pling at runtime—in other words, the behavior of architecture quanta as they interact
with one another to form workflows within a distributed architecture.
The nature of how services call one another creates difficult trade-off decisions
because it represents a multidimensional decision space, influenced by three inter‐
locking forces:
Communication
Refers to the type of connection synchronicity used: synchronous or
asynchronous.
Consistency
Describes whether the workflow communication requires atomicity or can utilize
eventual consistency.
Coordination
Describes whether the workflow utilizes an orchestrator or whether the services
communicate via choreography.
Communication
When two services communicate with each other, one of the fundamental questions
for an architect is whether that communication should be synchronous or asynchro‐
nous.
38
|
Chapter 2: Discerning Coupling in Software Architecture
Synchronous communication requires the requestor to wait for the response from the
receiver, as shown in Figure 2-11.
Figure 2-11. A synchronous call waits for a result from the receiver
The calling service makes a call (using one of a number of protocols that support syn‐
chronous calls, such as gRPC) and blocks (does no further processing) until the
receiver returns a value (or status indicating a state change or error condition).
Asynchronous communication occurs between two services when the caller posts a
message to the receiver (usually via a mechanism such as a message queue) and, once
the caller gets acknowledgment that the message will be processed, it returns to work.
If the request required a response value, the receiver can use a reply queue to (asyn‐
chronously) notify the caller of the result, which is illustrated in Figure 2-12.
Figure 2-12. Asynchronous communication allows parallel processing
The caller posts a message to a message queue and continues processing until notified
by the receiver that the requested information is available via return call. Generally,
architects use message queues (illustrated via the gray cylindrical tube in the top dia‐
gram in Figure 2-12) to implement asynchronous communication, but queues are
common and create noise on diagrams, so many architects leave them off, as shown
in the lower diagram. And, of course, architects can implement asynchronous com‐
munication without message queues by using a variety of libraries or frameworks.
Each diagram variety implies asynchronous messaging; the second provides visual
shorthand and less implementation detail.
Architecture (Quantum | Quanta)
|
39
Architects must consider significant trade-offs when choosing how services will com‐
municate. Decisions around communication affect synchronization, error handling,
transactionality, scalability, and performance. The remainder of this book delves into
many of these issues.
Consistency
Consistency refers to the strictness of transactional integrity that communication calls
must adhere to. Atomic transactions (all-or-nothing transactions requiring consis‐
tency during the processing of a request) lie on one side of the spectrum, whereas dif‐
ferent degrees of eventual consistency lie on the other side.
Transactionality—having several services participate in an all-or-nothing transaction
—is one of the most difficult problems to model in distibuted architectures, resulting
in the general advice to try to avoid cross-service transactions. We discuss consis‐
tency and the intersection of data and architecture in Chapters 6, 9, 10, and 12.
Coordination
Coordination refers to how much coordination the workflow modeled by the commu‐
nication requires. The two common generic patterns for microservices are orchestra‐
tion and choreography, which we describe in Chapter 11. Simple workflows—a single
service replying to a request—don’t require special consideration from this dimen‐
sion. However, as the complexity of the workflow grows, the greater the need for
coordination.
These three factors—communication, consistency, and coordination—all inform the
important decision an architect must make. Critically, however, architects cannot
make these choices in isolation; each option has a gravitation effect on the others. For
example, transactionality is easier in synchronous architectures with mediation,
whereas higher levels of scale are possible with eventually consistent asynchronous
choreographed systems.
Thinking about these forces as related to each other forms a three-dimensional space,
illustrated in Figure 2-13.
Each force in play during service communication appears as a dimension. For a par‐
ticular decision, an architect could graph the position in space representing the
strength of these forces.
40
|
Chapter 2: Discerning Coupling in Software Architecture
Figure 2-13. The dimensions of dynamic quantum coupling
When an architect can build a clear understanding of forces at play within a given
situation, it creates criteria for trade-off analysis. In the case of dynamic coupling,
Table 2-1 shows a framework for identifying fundamental pattern names based on the
eight possible combinations.
Table 2-1. The matrix of dimensional intersections for distributed architectures
Pattern name
Communication
Consistency
Coordination
Coupling
Epic Saga
Phone Tag Saga(sac)
Fairy Tale Saga(seo)
Time Travel Saga(sec)
Fantasy Fiction Saga(aao)
Horror Story(aac)
Parallel Saga(aeo)
Anthology Saga(aec)
synchronous
synchronous
synchronous
synchronous
asynchronous
asynchronous
asynchronous
asynchronous
atomic
atomic
eventual
eventual
atomic
atomic
eventual
eventual
orchestrated
choreographed
orchestrated
choreographed
orchestrated
choreographed
orchestrated
choreographed
very high
high
high
medium
high
medium
low
very low
(sao)
To fully understand this matrix, we must first investigate each of the dimensions indi‐
vidually. Therefore, the following chapters help you build context to understand the
individual trade-offs for communication, consistency, and coordination, then entan‐
gle them back together in Chapter 12.
In the remaining chapters in Part I, we focus on static coupling and understanding
the various dimensions at play in distributed architectures, including data ownership,
Architecture (Quantum | Quanta)
|
41
transactionality, and service granularity. In Part II, Putting Things Back Together, we
focus on dynamic coupling and understanding communication patterns in
microservices.
Sysops Squad Saga: Understanding Quanta
Tuesday, November 23, 14:32
Austen came to Addison’s office wearing an uncharacteristic cross expression. “Hey,
Addison, can I bother you for a minute?”
“Sure, what’s up?”
“I’ve been reading about this architecture quantum stuff, and I just…don’t…get…it!”
Addison laughed, “I know what you mean. I struggled with it when it was purely
abstract, but when you ground it in practical things, it turns out to be a useful set of
perspectives.”
“What do you mean?”
“Well,” said Addison, “the architecture quantum basically defines a DDD bounded context in architectural terms.”
“Why not just use bounded context, then?” asked Austen.
“Bounded context has a specific definition in DDD, and overloading it with stuff about architecture
just makes people constantly have to differentiate. They are similar, but not the same thing. The first
part about functional cohesion and independent deployment certainly matches a service based on
bounded context. But the architecture quantum definition goes further by identifying types of coupling—that’s where the static and dynamic stuff comes in.”
“What is that all about? Isn’t coupling just coupling? Why make the distinction?”
“It turns out that a bunch of different concerns revolve around the different types,” said Addison.
“Let’s take the static one first, which I like to think of as how things are wired together. Another way
to think about it: consider one of the services we’re building in our target architecture. What is all the
wiring required to bootstrap that service?”
“Well, it’s written in Java, using a Postgres database, and running in Docker—that’s it, right?”
“You’re missing a lot.” said Addison. “What if you had to build that service from scratch, assuming we
had nothing in place? It’s Java, but also using SpringBoot and, what, about 15 or 20 different frameworks and libraries?”
“That’s right, we can look in the Maven POM file to figure out all those dependencies. What else?”
“The idea behind static quantum coupling is the wiring required to function. We’re using events to
communicate between services—what about the event broker?”
42
|
Chapter 2: Discerning Coupling in Software Architecture
“But isn’t that the dynamic part?”
“Not the presence of the broker. If the service (or, more broadly, architecture quantum) I want to
bootstrap utilizes a message broker to function, the broker must be present. When the service calls
another service via the broker, we get into the dynamic side.”
“OK, that makes sense,” said Austen. “If I think about what it would take to bootstrap it from scratch,
that’s the static quantum coupling.”
“That’s right. And just that information is super useful. We recently built a diagram of the static quantum coupling for each of our services defensively.”
Austen laughed. “Defensively? What do you…”
“We were performing a reliability analysis to determine if I change this thing, what might break,
where thing could be anything in our architecture or operations. They’re trying to do risk mitigation
—if we change a service, they want to know what must be tested.”
“I see—that’s the static quantum coupling. I can see how that’s a useful view. It also shows how
teams might impact one another. That seems really useful. Is there a tool we can download that
figures that out for us?”
“Wouldn’t that be nice!” laughed Addison. “Unfortunately, no one with our unique mix of architecture has built and open sourced exactly the tool we want. However, some of the platform team is
working on a tool to automate it, necessarily customized to our architecture. They’re using the container manifests, POM files, NPM dependencies, and other dependency tools to build and maintain a
list of build dependencies. We have also instituted observability for all our services, so we now have
consistent log files about what systems call each other, when, and how often. They’re using that to
build a call graph to see how things are connected.”
“OK, so static coupling is how things are wired together. What about dynamic coupling?”
“Dynamic coupling concerns how quanta communicate with each other, particularly synchronous
versus asynchronous calls and their impact on operational architecture characteristics—things like
performance, scale, elasticity, reliability, and so on. Consider elasticity for a moment—remember the
difference between scalability and elasticity?”
Austen smirked. “I didn’t know there was going to be a test. Let’s see…scalability is the ability to support a large number of concurrent users; elasticity is the ability to support a burst of user requests in
a short time frame.”
“Correct! Gold star for you. OK, let’s think about elasticity. Suppose in our future state architecture we
have two services like Ticketing and Assignment, and the two types of calls. We carefully designed
our services to be highly statically decoupled from one another, so that they can be independently
elastic. That’s the other side effect of static coupling, by the way—it identifies the scope of things
like operational architecture characteristics. Let’s say that Ticketing is operating at 10 times the elastic scale of Assignment, and we need to make a call between them. If we make a synchronous call,
Sysops Squad Saga: Understanding Quanta
|
43
the whole workflow will bog down, as the caller waits for the slower service to process and return. If
on the other hand we make an asynchronous call, using the message queue as a buffer, we can
allow the two services to execute operationally independently, allowing the caller to add messages
to the queue and continue working, receiving notification when the workflow is complete.”
“Oh, I see, I see! The architecture quantum defines the scope of architecture characteristics—it’s
obvious how the static coupling can affect that. But I see now that, depending on the type of call
you make, you might temporarily couple two services together.”
“That’s right,” said Addison. “The architecture quanta can entangle one another temporarily, during
the course of a call, if the nature of the call ties things like performance, responsiveness, scale, and a
bunch of others.”
“OK, I think I understand what an architecture quantum is, and how the coupling definitions work.
But I’m never going to get that quantum/quanta thing straight!”
“Same for datum/data, but no one ever uses datum!” laughed Addison. “You’ll see a lot more of the
impact of dynamic coupling on workflows and transactional sagas as you keep digging into our
architecture.”
“I can’t wait!”
44
|
Chapter 2: Discerning Coupling in Software Architecture
CHAPTER 3
Architectural Modularity
Tuesday, September 21 09:33
It was the same conference room they had been in a hundred times before, but today
the atmosphere was different. Very different. As people gathered, no small talk was
exchanged. Only silence. The sort of dead silence that you could cut with a knife. Yes,
that was indeed an appropriate cliche given the topic of the meeting.
The business leaders and sponsors of the failing Sysops Squad ticketing application met
with the application architects, Addison and Austen, with the purpose of voicing their
concern and frustration about the inability of the IT department to fix the never-ending
issues associated with the trouble ticket application. “Without a working application,” they had said,
“we cannot possibly continue to support this business line.”
As the tense meeting ended, the business sponsors quietly filed out one by one, leaving Addison
and Austen alone in the conference room.
“That was a bad meeting,” said Addison. “I can’t believe they’re actually blaming us for all the issues
we’re currently facing with the trouble ticket application. This is a really bad situation.”
“Yeah, I know,” said Austen. “Especially the part about possibly closing down the product support
business line. We’ll be assigned to other projects, or worse, maybe even let go. Although I’d rather be
spending all of my time on the soccer field or on the slopes skiing in the winter, I really can’t afford
to lose this job.”
“Neither can I,” said Addison. “Besides, I really like the development team we have in place, and I’d
hate to see it broken up.”
“Me too,” said Austen. “I still think breaking apart the application would solve most of these issues.”
45
“I agree with you,” said Addison, “but how do we convince the business to spend more money and
time to refactor the architecture? You saw how they complained in the meeting about the amount
of money we’ve already spent applying patches here and there, only to create additional issues in
the process.”
“You’re right,” Austen said. “They would never agree to an expensive and time-consuming architecture migration effort at this point.”
“But if we both agree that we need to break apart the application to keep it alive, how in the world
are we going to convince the business and get the funding and time we need to completely
restructure the Sysops Squad application?” asked Addison.
“Beats me,” said Austen. “Let’s see if Logan is available to discuss this problem with us.”
Addison looked online and saw that Logan, the lead architect for Penultimate Electronics, was available. Addison sent a message explaining that they wanted to break apart the existing monolithic
application, but weren’t sure how to convince the business that this approach would work. Addison
explained in the message that they were in a real bind and could use some advice. Logan agreed to
meet with them and joined them in the conference room.
“What makes you so sure that breaking apart the Sysops Squad application will solve all of the
issues?” asked Logan.
“Because,” said Austen, “we’ve tried patching the code over and over, and it doesn’t seem to be working. We still have way too many issues.”
“You’re completely missing my point,” said Logan. “Let me ask you the question a different way. What
assurances do you have that breaking apart the system will accomplish anything more than just
spending more money and wasting more valuable time?”
“Well,” said Austen, “actually, we don’t.”
“Then how do you know breaking apart the application is the right approach?” asked Logan.
“We already told you,” said Austen, “because nothing else we try seems to work!”
“Sorry,” said Logan, “but you know as well as I do that’s not a reasonable justification for the business.
You’ll never get the funding you need with that kind of reason.”
“So, what would be a good business justification?” asked Addison. “How do we sell this approach to
the business and get the additional funding approved?”
“Well,” said Logan, “to build a good business case for something of this magnitude, you first need to
understand the benefits of architectural modularity, match those benefits to the issues you are facing with the current system, and finally analyze and document the trade-offs involved with breaking
apart the application.”
46
|
Chapter 3: Architectural Modularity
Businesses today face a torrent of change; market evolution seems to keep accelerat‐
ing at a blistering pace. Business drivers (such as mergers and acquisitions), increased
competition in the marketplace, increased consumer demand, and increased innova‐
tion (such as automation through machine learning and artificial intelligence) neces‐
sarily require changes to underlying computer systems. In many cases, those changes
in computer systems consequently necessitate changes to the underlying architectures
supporting them.
However, it’s not only business that’s undergoing constant and rapid change—it’s also
the technical environment in which those computer systems reside. Containerization,
the move to cloud-based infrastructure, the adoption of DevOps, and even new
advancements in continuous delivery pipelines all impact the underlying architecture
of those computer systems.
It’s difficult in today’s world to manage all of this constant and rapid change with
respect to software architecture. Software architecture is the foundational structure of
a system, and is therefore generally thought of as something that should remain sta‐
ble and not undergo frequent change, similar to the underlying structural aspects of a
large building or skyscraper. However, unlike the structural architecture of a building,
software architecture must constantly change and adapt to meet the new demands of
today’s business and technology environment.
Consider the increased number of mergers and acquisitions happening in today’s
marketplace. When one company acquires another, not only does it acquire the phys‐
ical aspects of a company (such as people, buildings, inventory, and so on) but also
more customers. Can the existing systems in either company scale to meet the
increase in user volume as a result of the merger or acquisition? Scalability is a big
part of mergers and acquisitions, as is agility and extensibility, all of which are archi‐
tectural concerns.
Large monolithic (single deployment) systems generally do not provide the level of
scalability, agility, and extensibility required to support most mergers and acquisi‐
tions. The capacity for additional machine resources (threads, memory, and CPU)
fills up very quickly. To illustrate this point, consider the water glass shown in Figure
3-1. The glass represents the server (or virtual machine), and the water represents the
application. As monolithic applications grow to handle increased consumer demand
and user load (whether from mergers, acquisitions, or company growth), they begin
to consume more and more resources. As more water is added to the glass (represent‐
ing the growing monolithic application), the glass begins to fill up. Adding another
glass (represented as another server or virtual machine) does nothing, because the
new glass would contain the same amount of water as the first one.
Architectural Modularity
|
47
Figure 3-1. A full glass representing a large monolithic application close to capacity
One aspect of architectural modularity is breaking large monolithic applications into
separate and smaller parts to provide more capacity for further scalability and
growth, while at the same time facilitating constant and rapid change. In turn, these
capabilities can help achieve a company’s strategic goals.
By adding another empty glass to our water glass example and breaking the water
(application) into two separate parts, half the water can now be poured into the new
empty glass, providing 50% more capacity, as shown in Figure 3-2. The water glass
analogy is a great way of explaining architectural modularity (the breaking up of
monolithic applications) to business stakeholders and C-level executives, who will
inevitably be paying for the architecture-refactoring effort.
Figure 3-2. Two half-full glasses representing an application broken apart with plenty of
capacity for growth
48
|
Chapter 3: Architectural Modularity
Increased scalability is only one benefit of architectural modularity. Another impor‐
tant benefit is agility, the ability to respond quickly to change. An article from Forbes
in January 2020 by David Benjamin and David Komlos stated the following:
There is one thing that will separate the pack into winners and losers: the on-demand
capability to make bold and decisive course-corrections that are executed effectively
and with urgency.
Businesses must be agile in order to survive in today’s world. However, while business
stakeholders may be able to make quick decisions and change direction quickly, the
company’s technology staff may not be able to implement those new directives fast
enough to make a difference. Enabling technology to move as fast as the business (or,
conversely, preventing technology from slowing the business) requires a certain level
of architectural agility.
Modularity Drivers
Architects shouldn’t break a system into smaller parts unless clear business drivers
exist. The primary business drivers for breaking applications into smaller parts
include speed-to-market (sometimes called time-to-market) and achieving a level of
competitive advantage in the marketplace.
Speed-to-market is achieved through architectural agility—the ability to respond
quickly to change. Agility is a compound architectural characteristic made up of
many other architecture characteristics, including maintainability, testability, and
deployability.
Competitive advantage is achieved through speed-to-market combined with scalabil‐
ity and overall application availability and fault tolerance. The better a company does,
the more it grows, hence the need for more scalability to support increased user activ‐
ity. Fault tolerance, the ability of an application to fail and continue to operate, is nec‐
essary to ensure that as parts of the application fail, other parts are still able to
function as normal, minimizing the overall impact to the end user. Figure 3-3 illus‐
trates the relationship between the technical drivers and the resulting business drivers
for modularity (enclosed within boxes).
Businesses must be agile to survive in today’s fast-paced and ever-changing volatile
market, meaning the underlying architectures must be agile as well. As illustrated in
Figure 3-3, the five key architectural characteristics to support agility, speed-tomarket, and, ultimately, competitive advantage in today’s marketplace are availability
(fault tolerance), scalability, deployability, testability, maintainability.
Modularity Drivers
|
49
Figure 3-3. The drivers for modularity and the relationships among them
Note that architectural modularity does not always have to translate to a distributed
architecture. Maintainability, testability, and deployability (defined in the following
sections) can also be achieved through monolithic architectures such as a modular
monolith or even a microkernel architecture (see Appendix B for a list of references
providing more information about these architecture styles). Both of these architec‐
ture styles offer a level of architectural modularity based on the way the components
are structured. For example, with a modular monolith, components are grouped into
well-formed domains, providing for what is known as a domain partitioned architec‐
ture (see Fundamentals of Software Architecture, Chapter 8, page 103). With the
microkernel architecture, functionality is partitioned into separate plug-in compo‐
nents, allowing for a much smaller testing and deployment scope.
Maintainability
Maintainability is about the ease of adding, changing, or removing features, as well as
applying internal changes such as maintenance patches, framework upgrades, thirdparty upgrades, and so on. As with most composite architecture characteristics,
maintainability is hard to define objectively. Alexander von Zitzewitz, software archi‐
tect and founder of hello2morrow, wrote an article about a new metric for objectively
defining the maintainability level of an application. While von Zitzewitz’s maintaina‐
bility metric is fairly complicated and involves lots of factors, its initial form is as
follows:
50
|
Chapter 3: Architectural Modularity
ML = 100 *
k
∑ ci
i=1
where ML is the maintainability level of the overall system (percentage from 0% to
100%), k is the total number of logical components in the system, and ci is the cou‐
pling level for any given component, with a special focus on incoming coupling lev‐
els. This equation basically demonstrates that the higher the incoming coupling level
between components, the lower the overall maintainability level of the codebase.
Putting aside complicated mathematics, some of the typical metrics used for deter‐
mining the relative maintainability of an application based on components (the archi‐
tectural building blocks of an application) include the following:
Component coupling
The degree and manner to which components know about one another
Component cohesion
The degree and manner to which the operations of a component interrelate
Cyclomatic complexity
The overall level of indirection and nesting within a component
Component size
The number of aggregated statements of code within a component
Technical versus domain partitioning
Components aligned by technical usage or by domain purpose—see Appendix A
Within the context of architecture, we are defining a component as an architectural
building block of the application that does some sort of business or infrastructure
function, usually manifested through a package structure (Java), namespace (C#), or
physical grouping of files (classes) within some sort of directory structure. For exam‐
ple, the component Order History might be implemented through a set of class files
located in the namespace app.business.order.history.
Large monolithic architectures generally have low levels of maintainability due to the
technical partitioning of functionality into layers, the tight coupling between compo‐
nents, and weak component cohesion from a domain perspective. For example, con‐
sider a new requirement within a traditional monolithic layered architecture to add
an expiration date to items contained in a customer’s wish list (items in a list to
maybe purchase at a later time). Notice in Figure 3-4 that the change scope of the new
requirement is at an application level since the change is propagated to all of the layers
within the application.
Modularity Drivers
|
51
Figure 3-4. With monolithic layered architectures, change is at an application level
Depending on the team structure, implementing this simple change to add an expira‐
tion date to wish list items in a monolithic layered architecture could possibly require
the coordination of at least three teams:
• A member from the user interface team would be needed to add the new expiry
field to the screen.
• A member from the backend team would be needed to add business rules associ‐
ated with the expiry date and change contracts to add the new expiry field.
• A member from the database team would be needed to change the table schema
to add the new expiry column in the Wishlist table.
Since the Wishlist domain is spread throughout the entire architecture, it becomes
harder to maintain a particular domain or subdomain (such as Wishlist). Modular
architectures, on the other hand, partition domains and subdomains into smaller,
separately deployed units of software, thereby making it easier to modify a domain or
subdomain. Notice that with a distributed service-based architecture, as shown in
Figure 3-5, the change scope of the new requirement is at a domain level within a par‐
ticular domain service, making it easier to isolate the specific deployment unit requir‐
ing the change.
Moving to even more architectural modularity such as a microservices architecture,
as illustrated in Figure 3-6, places the new requirement at a function-level change
52
|
Chapter 3: Architectural Modularity
scope, isolating the change to a specific service responsible for the wish list
functionality.
Figure 3-5. With service-based architectures, change is at a domain level
Figure 3-6. With microservices architectures, change is at a function level
Modularity Drivers
|
53
These three progressions toward modularity demonstrate that as the level of architec‐
tural modularity increases, so does maintainability, making it easier to add, change,
or remove functionality.
Testability
Testability is defined as the ease of testing (usually implemented through automated
tests) as well as the completeness of testing. Testability is an essential ingredient for
architectural agility. Large monolithic architecture styles like the layered architecture
support relatively low levels of testability (and hence agility) due to the difficulty in
achieving full and complete regression testing of all features within the large deploy‐
ment unit. Even if a monolithic application did have a suite of full regression tests,
imagine the frustration of having to execute hundreds or even thousands of unit tests
for a simple code change. Not only would it take a long time to execute all of the tests,
but the poor developer would be stuck researching why dozens of tests failed when in
fact the failed tests have nothing to do with the change.
Architectural modularity—the breaking apart of applications into smaller deploy‐
ment units—significantly reduces the overall testing scope for changes made to a ser‐
vice, allowing for better completeness of testing as well as ease of testing. Not only
does modularity result in smaller, more targeted test suites, but maintaining the unit
tests becomes easier as well.
While architectural modularity generally improves testability, it can sometimes lead
to the same problems that exist with monolithic, single-deployment applications. For
example, consider an application that was broken into three smaller self-contained
deployment units (services), as depicted in Figure 3-7.
Making a change to Service A limits the testing scope to only that service, since Ser‐
vice B and Service C are not coupled to Service A. However, as communication
increases among these services, as shown at the bottom of Figure 3-7, testability
declines rapidly because the testing scope for a change to Service A now includes Ser‐
vice B and Service C, therefore impacting both the ease of testing and the complete‐
ness of testing.
54
| Chapter 3: Architectural Modularity
Figure 3-7. Testing scope is increased as services communicate with one another
Deployability
Deployability is not only about the ease of deployment—it is also about the frequency
of deployment and the overall risk of deployment. To support agility and respond
quickly to change, applications must support all three of these factors. Deploying
software every two weeks (or more) not only increases the overall risk of deployment
(due to grouping multiple changes together), but in most cases unnecessarily delays
new features or bug fixes that are ready to be pushed out to customers. Of course,
deployment frequency must be balanced with the customer’s (or end user’s) ability to
be able to absorb changes quickly.
Monolithic architectures generally support low levels of deployability due to the
amount of ceremony involved in deploying the application (such as code freezes,
mock deployments, and so on), the increased risk that something else might break
once new features or bug fixes are deployed, and a long time frame between deploy‐
ments (weeks to months). Applications having a certain level of architectural modu‐
larity in terms of separately deployed units of software have less deployment
ceremony, less risk of deployment, and can be deployed more frequently than a large,
single monolithic application.
Modularity Drivers
|
55
Like testability, deployability is also negatively impacted as services become smaller
and communicate more with each other to complete a business transaction. Deploy‐
ment risk is increased, and it becomes more difficult to deploy a simple change for
fear of breaking other services. To quote software architect Matt Stine in his article on
orchestrating microservices:
If your microservices must be deployed as a complete set in a specific order, please put
them back in a monolith and save yourself some pain.
This scenario leads to what is commonly referred to as the “big ball of distributed
mud,” where very few (if any) of the benefits of architectural modularity are realized.
Scalability
Scalability is defined as the ability of a system to remain responsive as user load grad‐
ually increases over time. Related to scalability is elasticity, which is defined as the
ability of a system to remain responsive during significantly high instantaneous and
erratic spikes in user load. Figure 3-8 illustrates the differences between scalability
and elasticity.
Figure 3-8. Scalability is different from elasticity
While both of these architectural characteristics include responsiveness as a function
of the number of concurrent requests (or users in the system), they are handled dif‐
ferently from an architectural and implementation standpoint. Scalability generally
occurs over a longer period of time as a function of normal company growth, whereas
elasticity is the immediate response to a spike in user load.
A great example to further illustrate the difference is that of a concert-ticketing
system. Between major concert events, there is usually a fairly light concurrent user
load. However, the minute tickets go on sale for a popular concert, concurrent user
load significantly spikes. The system may go from 20 concurrent users to 3,000 con‐
current users in a matter of seconds. To maintain responsiveness, the system must
have the capacity to handle the high peaks in user load, and also have the ability to
instantaneously start up additional services to handle the spike in traffic. Elasticity
relies on services having a very small mean time to startup (MTTS), which is achieved
56
|
Chapter 3: Architectural Modularity
architecturally by having very small, fine-grained services. With an appropriate archi‐
tectural solution in place, MTTS (and hence elasticity) can then be further managed
through design-time techniques such as small lightweight platforms and runtime
environments.
Although both scalability and elasticity improve with finer-grained services, elasticity
is more a function of granularity (the size of a deployment unit), whereas scalability is
more a function of modularity (the breaking apart of applications into separate
deployment units). Consider the traditional layered architecture, service-based archi‐
tecture, and microservices architecture styles and their corresponding star ratings for
scalability and elasticity, as illustrated in Figure 3-9 (the details of these architecture
styles and their corresponding star ratings can be found in our previous book, Funda‐
mentals of Software Architecture. Note that one star means that the capability is not
well supported by the architecture style, whereas five stars means that capability is a
major feature of the architecture style and is well supported.
Notice that scalability and elasticity rate relatively low with the monolithic layered
architecture. Large monolithic layered architectures are both difficult and expensive
to scale because all of the application functionality must scale to the same degree
(application-level scalability and poor MTTS). This can become particularly costly in
cloud-based infrastructures. However, with service-based architecture, notice that
scalability improves, but not as much as elasticity. This is because domain services in
a service-based architecture are coarse grained and usually contain the entire domain
in one deployment unit (such as order processing or warehouse management), and
generally have too long of a mean time to startup (MTTS) to respond fast enough to
immediate demand for elasticity due to their large size (domain-level scalability and
fair MTTS). Notice that with microservices, both scalability and elasticity are maxi‐
mized because of the small, single-purpose, fine-grained nature of each separately
deployed service (function-level scalability and excellent MTTS).
Like testability and deployability, the more services communicate with one other to
complete a single business transaction, the greater the negative impact on scalability
and elasticity. For this reason, it is important to keep synchronous communication
among services to a minimum when requiring high levels of scalability and elasticity.
Modularity Drivers
|
57
Figure 3-9. Scalability and elasticity improve with modularity
Availability/Fault Tolerance
Like many architecture characteristics, fault tolerance has varying definitions. Within
the context of architectural modularity, we define fault tolerance as the ability for
some parts of the system to remain responsive and available as other parts of the sys‐
tem fail. For example, if a fatal error (such as an out-of-memory condition) in the
payment-processing portion of a retail application occurs, the users of the system
should still be able to search for items and place orders, even though the payment
processing is unavailable.
All monolithic systems suffer from low levels of fault tolerance. While fault tolerance
can be somewhat mitigated in a monolithic system by having multiple instances of
the entire application load balanced, this technique is both expensive and ineffective.
If the fault is due to a programming bug, that bug will exist in both instances, there‐
fore potentially bringing down both instances.
Architectural modularity is essential to achieving domain-level and function-level
fault tolerance in a system. By breaking apart the system into multiple deployment
58
|
Chapter 3: Architectural Modularity
units, catastrophic failure is isolated to only that deployment unit, thereby allowing
the rest of the system to function normally. There is a caveat to this, however: if other
services are synchronously dependent on a service that is failing, fault tolerance is not
achieved. This is one of the reasons asynchronous communication between services
is essential for maintaining a good level of fault tolerance in a distributed system.
Sysops Squad Saga: Creating a Business Case
Thursday, September 30, 12:01
Armed with a better understanding of what is meant by architectural modularity and the
corresponding drivers for breaking apart a system, Addison and Austen met to discuss
the Sysops Squad issues and try to match them to modularity drivers in order to build a
solid business justification to present to the business sponsors.
“Let’s take each of the issues we are facing and see if we can match them to some of the
modularity drivers,” said Addison. “That way, we can demonstrate to the business that
breaking apart the application will in fact address the issues we are facing.”
“Good idea,” said Austen. “Let’s start with the first issue they talked about in the meeting—change.
We cannot seem to effectively apply changes to the existing monolithic system without something
else breaking. Also, changes take way too long, and testing the changes is a real pain.”
“And the developers are constantly complaining that the codebase is too large, and it’s difficult to
find the right place to apply changes to new features or bug fixes,” said Addison.
“OK,” said Austen, “so clearly, overall maintainability is a key issue here.”
“Right,” said Addison. “So, by breaking apart the application, it would not only decouple the code,
but it would isolate and partition the functionality into separately deployed services, making it easier
for developers to apply changes.”
“Testability is another key characteristic related to this problem, but we have that covered already
because of all our automated unit tests,” said Austen.
“Actually, it’s not,” replied Addison. “Take a look at this.”
Addison showed Austen that over 30% of the test cases are commented out or obsolete, and there
are missing test cases for some of the critical workflow parts of the system. Addison also explained
that the developers were continually complaining that the entire unit test suite had to be run for any
change (big or small), which not only took a long time, but developers were faced with having to fix
issues not related to their change. This was one of the reasons it was taking so long to apply even
the simplest of changes.
“Testability is about the ease of testing, but also the completeness of testing,” said Addison. “We
have neither. By breaking apart the application, we can significantly reduce the scope of testing for
Sysops Squad Saga: Creating a Business Case
|
59
changes made to the application, group relevant automated unit tests together, and get better
completeness of testing—hence fewer bugs.”
“The same is true with deployability,” continued Addison. “Because we have a monolithic application, we have to deploy the entire system, even for a small bug fix. Because our deployment risk is so
high, Parker insists on doing production releases on a monthly basis. What Parker doesn’t understand is that by doing so, we pile multiple changes onto every release, some of which haven’t even
been tested in conjunction with each other.”
“I agree,” said Austen, “and besides, the mock deployments and code freezes we do for each release
take up valuable time—time we don’t have. However, what we’re talking about here is not an architecture issue, but purely a deployment pipeline issue.”
“I disagree,” said Addison. “It’s definitely architecture related as well. Think about it for a minute, Austen. If we broke the system into separately deployed services, then a change for any given service
would be scoped to that service only. For example, let’s say we make yet another change to the
ticket assignment process. If that process was a separate service, not only would the testing scope
be reduced, but we would significantly reduce the deployment risk. That means we could deploy
more frequently with much less ceremony, as well as significantly reduce the number of bugs.”
“I see what you mean,” said Austen, “and while I agree with you, I still maintain that at some point we
will have to modify our current deployment pipeline as well.”
Satisfied that breaking apart the Sysops Squad application and moving to a distributed architecture
would address the change issues, Addison and Austen moved on to the other business sponsor
concerns.
“OK,” said Addison, “the other big thing the business sponsors complained about in the meeting was
overall customer satisfaction. Sometimes the system isn’t available, the system seems to crash at certain times during the day, and we’ve experienced too many lost tickets and ticket routing issues. It’s
no wonder customers are starting to cancel their support plans.”
“Hold on,” said Austen. “I have some latest metrics here that show it’s not the core ticketing functionality that keeps bringing the system down, but the customer survey functionality and reporting.”
“This is excellent news,” said Addison. “So by breaking apart that functionality of the system into separate services, we can isolate those faults, keeping the core ticketing functionality operational. That’s
a good justification in and of itself!”
“Exactly,” said Austen. “So, we are in agreement then that overall availability through fault tolerance
will address the application not always being available for the customers since they only interact
with the ticketing portion of the system.”
“But what about the system freezing up?” asked Addison. “How do we justify that part with breaking
up the application?”
60
|
Chapter 3: Architectural Modularity
“It just so happens I asked Sydney from the Sysops Squad development team to run some analysis
for me regarding exactly that issue,” said Austen. “It turns out that it is a combination of two things.
First, whenever we have more than 25 customers creating tickets at the same time, the system
freezes. But, check this out—whenever they run the operational reports during the day when customers are entering problem tickets, the system also freezes up.”
“So,” said Addison, “it appears we have both a scalability and a database load issue here.”
“Exactly!” Austen said. “And get this—by breaking up the application and the monolithic database,
we can segregate reporting into its own system and also provide the added scalability for the
customer-facing ticketing functionality.”
Satisfied that they had a good business case to present to the business sponsors and confident that
this was the right approach for saving this business line, Addison created an Architecture Decision
Record (ADR) for the decision to break apart the system and create a corresponding business case
presentation for the business sponsors.
ADR: Migrate Sysops Squad Application to a Distributed Architecture
Context
The Sysops Squad is currently a monolithic problem ticket application that supports many
different business functions related to problem tickets, including customer registration,
problem ticket entry and processing, operations and analytical reporting, billing and payment processing, and various administrative maintenance functions. The current application has numerous issues involving scalability, availability, and maintainability.
Decision
We will migrate the existing monolithic Sysops Squad application to a distributed architecture. Moving to a distributed architecture will accomplish the following:
• Make the core ticketing functionality more available for our external customers,
therefore providing better fault tolerance
• Provide better scalability for customer growth and ticket creation, resolving the frequent application freeze-ups we’ve been experiencing
• Separate the reporting functionality and reporting load on the database, resolving
the frequent application freeze-ups we’ve been experiencing
• Allow teams to implement new features and fix bugs much faster than with the current monolithic application, therefore providing for better overall agility
• Reduce the amount of bugs introduced into the system when changes occur, therefore providing better testability
• Allow us to deploy new features and bug fixes at a much faster rate (weekly or even
daily), therefore providing better deployability
Sysops Squad Saga: Creating a Business Case
|
61
Consequences
The migration effort will cause delays for new features being introduced since most of the
developers will be needed for the architecture migration.
The migration effort will incur additional cost (cost estimates to be determined).
Until the existing deployment pipeline is modified, release engineers will have to manage
the release and monitoring of multiple deployment units.
The migration effort will require us to break apart the monolithic database.
Addison and Austen met with the business sponsors for the Sysops Squad problem ticketing system
and presented their case in a clear and concise manner. The business sponsors were pleased with
the presentation and agreed with the approach, informing Addison and Austen to move forward
with the migration.
62
|
Chapter 3: Architectural Modularity
CHAPTER 4
Architectural Decomposition
Monday, October 4, 10:04
Now that Addison and Austen had the go-ahead to move to a distributed architecture
and break apart the monolithic Sysops Squad application, they needed to determine the
best approach for how to get started.
“The application is so big I don’t even know where to start. It’s as big as an elephant!”
exclaimed Addison.
“Well,” said Austen. “How do you eat an elephant?”
“Ha, I’ve heard that joke before, Austen. One bite at a time, of course!” laughed Addison.
“Exactly. So let’s use the same principle with the Sysops Squad application,” said Austen. “Why don’t
we just start breaking it apart, one bite at a time? Remember how I said reporting was one of the
things causing the application to freeze up? Maybe we should start there.”
“That might be a good start,” said Addison, “but what about the data? Just making reporting a separate service doesn’t solve the problem. We’d need to break apart the data as well, or even create a
separate reporting database with data pumps to feed it. I think that’s too big of a bite to take starting
out.”
“You’re right,” said Austen. “Hey, what about the knowledge base functionality? That’s fairly standalone and might be easier to extract.”
“That’s true. And what about the survey functionality? That should be easy to separate out as well,”
said Addison. “The problem is, I can’t help feeling like we should be tackling this with more of a
methodical approach rather than just eating the elephant bite by bite.”
“Maybe Logan can give us some advice,” said Austen.
63
Addison and Austen met with Logan to discuss some of the approaches they were considering for
how to break apart the application. They explained to Logan that they wanted to start with the
knowledge base and survey functionality but weren’t sure what to do after that.
“The approach you’re suggesting,” said Logan, “is what is known as the Elephant Migration AntiPattern. Eating the elephant one bite at a time may seem like a good approach at the start, but in
most cases it leads to an unstructured approach that results in a big ball of distributed mud, what
some people also call a distributed monolith. I would not recommend that approach.”
“So, what other approaches exist? Are there patterns we can use to break apart the application?”
asked Addison.
“You need to take a holistic view of the application and apply either tactical forking or componentbased decomposition,” said Logan. “Those are the two most effective approaches I know of.”
Addison and Austen looked at Logan. “But how do we know which one to use?”
Whereas architectural modularity describes the why for breaking apart a monolithic
application, architectural decomposition describes the how. Breaking apart large,
complex monolithic applications can be a complex and time-consuming undertaking,
and it’s important to know whether it is even feasible to begin such an effort and how
to approach it.
Component-based decomposition and tactical forking are two common approaches
for breaking apart monolithic applications. Component-based decomposition is an
extraction approach that applies various refactoring patterns for refining and extract‐
ing components (the logical building blocks of an application) to form a distributed
architecture in an incremental and controlled fashion. The tactical forking approach
involves making replicas of an application and chipping away at the unwanted parts
to form services, similar to the way a sculptor creates a beautiful work of art from a
block of granite or marble.
Which approach is most effective? The answer to this question is, of course, it
depends. One of the main factors in selecting a decomposition approach is how well
the existing monolithic application code is structured. Do clear components and
component boundaries exist within the codebase, or is the codebase largely an
unstructured big ball of mud?
As the flowchart in Figure 4-1 illustrates, the first step in an architecture decomposi‐
tion effort is to first determine whether the codebase is even decomposable. We cover
this topic in detail in the next section. If the codebase is decomposable, the next step
is to determine if the source code is largely an unstructured mess with no clearly
definable components. If that’s the case, then tactical forking (see “Tactical Forking”
on page 73) is probably the right approach. However, if the source code files are
structured in a way that combines like functionality within well-defined (or even
64
|
Chapter 4: Architectural Decomposition
loosely defined) components, then a component-based decomposition approach (see
“Component-Based Decomposition” on page 71) is the way to go.
Figure 4-1. The decision tree for selecting a decomposition approach
We describe both of these approaches in this chapter, and then devote an entire chap‐
ter (Chapter 5) to describing each of the component-based decomposition patterns in
detail.
Is the Codebase Decomposable?
What happens when a codebase lacks internal structure? Can it even be decomposed?
Such software has a colloquial name—the Big Ball of Mud Anti-Pattern, coined by
Brian Foote in a same-named essay in 1999. For example, a complex web application
with event handlers wired directly to database calls and no modularity can be consid‐
ered a Big Ball of Mud architecture. Generally, architects don’t spend much time cre‐
ating patterns for these kinds of systems; software architecture concerns internal
structure, and these systems lack that defining feature.
Is the Codebase Decomposable?
|
65
Unfortunately, without careful governance, many software systems degrade into big
balls of mud, leaving it to subsequent architects (or perhaps a despised former self) to
repair. Step one in any architecture restructuring exercise requires an architect to
determine a plan for the restructuring, which in turn requires the architect to under‐
stand the internal structure. The key question the architect must answer becomes is
this codebase salvageable? In other words, is it a candidate for decomposition pat‐
terns, or is another approach more appropriate?
No single measure will determine whether a codebase has reasonable internal struc‐
ture—that evaluation falls to one or more architects to determine. However, archi‐
tects do have tools to help determine macro characteristics of a codebase, particularly
coupling metrics, to help evaluate internal structure.
Afferent and Efferent Coupling
In 1979, Edward Yourdon and Larry Constantine published Structured Design: Fun‐
damentals of a Discipline of Computer Program and Systems Design (Yourdon), defin‐
ing many core concepts, including the metrics afferent and efferent coupling. Afferent
coupling measures the number of incoming connections to a code artifact (compo‐
nent, class, function, and so on). Efferent coupling measures the outgoing connections
to other code artifacts.
Note the value of just these two measures when changing the structure of a system.
For example, when deconstructing a monolith into a distributed architecture, an
architect will find shared classes such as Address. When building a monolith, it is
common and encouraged for developers to reuse core concepts such as Address, but
when pulling the monolith apart, now the architect must determine how many other
parts of the system use this shared asset.
Virtually every platform has tools that allow architects to analyze the coupling char‐
acteristics of code in order to assist in restructuring, migrating, or understanding a
codebase. Many tools exist for various platforms that provide a matrix view of class
and/or component relationships, as illustrated in Figure 4-2.
In this example, the Eclipse plug-in provides a visualization of the output of JDepend,
which includes coupling analysis per package, along with some aggregate metrics
highlighted in the next section.
66
|
Chapter 4: Architectural Decomposition
Figure 4-2. JDepend in Eclipse analysis view of coupling relationships
Abstractness and Instability
Robert Martin, a well-known figure in the software architecture world, created some
derived metrics for a C++ book in the late 1990s that are applicable to any objectoriented language. These metrics—abstractness and instability—measure the balance
of the internal characteristics of a codebase.
Abstractness is the ratio of abstract artifacts (abstract classes, interfaces, and so on) to
concrete artifacts (implementation classes). It represents a measure of abstract versus
implementation. Abstract elements are features of a codebase that allow developers to
understand the overall function better. For example, a codebase consisting of a single
main() method and 10,000 lines of code would score nearly zero on this metric and
be quite hard to understand.
The formula for abstractness appears in Equation 4-1.
Equation 4-1. Abstractness
A=
∑ ma
∑ mc + ∑ ma
Is the Codebase Decomposable?
|
67
In the equation, ma represents abstract elements (interfaces or abstract classes) within
the codebase, and mc represents concrete elements. Architects calculate abstractness
by calculating the ratio of the sum of abstract artifacts to the sum of the concrete
ones.
Another derived metric, instability, is the ratio of efferent coupling to the sum of both
efferent and afferent coupling, shown in Equation 4-2.
Equation 4-2. Instability
I=
Ce
Ce + Ca
In the equation, Ce represents efferent (or outgoing) coupling, and Ca represents
afferent (or incoming) coupling.
The instability metric determines the volatility of a codebase. A codebase that exhibits
high degrees of instability breaks more easily when changed because of high cou‐
pling. Consider two scenarios, each with Ca of 2. For the first scenario, Ce = 0, yield‐
ing an instability score of zero. In the other scenario, Ce = 3, yielding an instability
score of 3/5. Thus, the measure of instability for a component reflects how many
potential changes might be forced by changes to related components. A component
with an instability value near one is highly unstable, a value close to zero may be
either stable or rigid: it is stable if the module or component contains mostly abstract
elements, and rigid if it comprises mostly concrete elements. However, the trade-off
for high stability is lack of reuse—if every component is self contained, duplication is
likely.
A component with an I value close to 1, we can agree, is highly instable. However, a
component with a value of I close to 0 may be either stable or rigid. However, if it
contains mostly concrete elements, then it is rigid.
Thus, in general, it is important to look at the value of I and A together rather than in
isolation. Hence the reason to consider the main sequence presented on the next
page.
68
| Chapter 4: Architectural Decomposition
Distance from the Main Sequence
One of the few holistic metrics architects have for architectural structure is distance
from the main sequence, a derived metric based on instability and abstractness, shown
in Equation 4-3.
Equation 4-3. Distance from the main sequence
D= A+I−1
In the equation, A = abstractness and I = instability.
The distance-from-the-main-sequence metric imagines an ideal relationship between
abstractness and instability; components that fall near this idealized line exhibit a
healthy mixture of these two competing concerns. For example, graphing a particular
component allows developers to calculate the distance-from-the-main-sequence met‐
ric, illustrated in Figure 4-3.
Figure 4-3. Normalized distance from the main sequence for a particular component
Developers graph the candidate component, then measure the distance from the ide‐
alized line. The closer to the line, the better balanced the component. Components
that fall too far into the upper-right corner enter into what architects call the zone of
uselessness: code that is too abstract becomes difficult to use. Conversely, code that
falls into the lower-left corner enter the zone of pain: code with too much implemen‐
tation and not enough abstraction becomes brittle and hard to maintain, illustrated in
Figure 4-4.
Is the Codebase Decomposable?
|
69
Tools exist in many platforms to provide these measures, which assist architects when
analyzing codebases because of unfamiliarity, migration, or technical debt
assessment.
What does the distance-from-the-main-sequence metric tell architects looking to
restructure applications? Just as in construction projects, moving a large structure
that has a poor foundation presents risks. Similarly, if an architect aspires to restruc‐
ture an application, improving the internal structure will make it easier to move the
entity.
Figure 4-4. Zones of uselessness and pain
This metric also provides a good clue as to the balance of the internal structure. If an
architect evaluates a codebase where many of the components fall into either the
zones of uselessness or pain, perhaps it is not a good use of time to try to shore up the
internal structure to the point where it can be repaired.
Following the flowchart in Figure 4-1, once an architect decides that the codebase is
decomposable, the next step is to determine what approach to take to decompose the
application. The following sections describe the two approaches for decomposing an
application: component-based decomposition and tactical forking.
70
|
Chapter 4: Architectural Decomposition
Component-Based Decomposition
It has been our experience that most of the difficulty and complexity involved with
migrating monolithic applications to highly distributed architecture like microservi‐
ces comes from poorly defined architectural components. Here we define a compo‐
nent as a building block of the application that has a well-defined role and
responsibility in the system and a well-defined set of operations. Components in
most applications are manifested through namespaces or directory structures and are
implemented through component files (or source files). For example, in Figure 4-5,
the directory structure penultimate/ss/ticket/assign would represent a component
called Ticket Assign with the namespace penultimate.ss.ticket.assign.
Figure 4-5. The directory structure of a codebase becomes the namespace of the
component
When breaking monolithic applications into distributed architec‐
tures, build services from components, not individual classes.
Throughout many collective years of migrating monolithic applications to distributed
architectures (such as microservices), we’ve developed a set of component-based
decomposition patterns described in Chapter 5 that help prepare a monolithic appli‐
cation for migration. These patterns involve the refactoring of source code to arrive at
a set of well-defined components that can eventually become services, easing the
effort needed to migrate applications to distributed architectures.
Component-Based Decomposition
|
71
These component-based decomposition patterns essentially enable the migration of a
monolithic architecture to a service-based architecture, which is defined in Chapter 2
and described in more detail in Fundamentals of Software Architecture. Service-based
architecture is a hybrid of the microservices architecture style where an application is
broken into domain services, which are coarse-grained, separately deployed services
containing all of the business logic for a particular domain.
Moving to a service-based architecture is suitable as a final target or as a steppingstone to microservices:
• As a stepping-stone, it allows an architect to determine which domains require
further levels of granularity into microservices and which ones can remain as
coarse-grained domain services (this decision is discussed in detail in Chapter 7).
• Service-based architecture does not require the database to be broken apart,
therefore allowing architects to focus on the domain and functional partitioning
prior to tackling database decomposition (discussed in detail in Chapter 6).
• Service-based architecture does not require any operational automation or con‐
tainerization. Each domain service can be deployed using the same deployment
artifact as the original application (such as an EAR file, WAR file, Assembly, and
so on).
• The move to service-based architecture is a technical one, meaning it generally
doesn’t involve business stakeholders and doesn’t require any change to the orga‐
nization structure of the IT department nor the testing and deployment
environments.
When migrating monolithic applications to microservices, con‐
sider moving to a service-based architecture first as a steppingstone to microservices.
But what if the codebase is an unstructured big ball of mud and doesn’t contain very
many observable components? That’s where tactical forking comes in.
72
| Chapter 4: Architectural Decomposition
Tactical Forking
The tactical forking pattern was named by Fausto De La Torre as a pragmatic
approach to restructuring architectures that are basically big balls of mud.
Generally, when architects think about restructuring a codebase, they think of
extracting pieces, as illustrated in Figure 4-6.
Figure 4-6. Extracting a part of a system
However, another way to think of isolating one part of a system involves deleting the
parts no longer needed, as illustrated in Figure 4-7.
Figure 4-7. Deleting what’s not wanted is another way to isolate parts of a system
In Figure 4-6, developers have to constantly deal with the exuberant strands of cou‐
pling that define this architecture; as they extract pieces, they discover that more and
more of the monolith must come along because of dependencies. In Figure 4-7,
developers delete what code isn’t needed, but the dependencies remain, avoiding the
constant unraveling effect of extraction.
Tactical Forking
|
73
The difference between extraction and deletion inspires the tactical forking pattern.
For this decomposition approach, the system starts as a single monolithic application,
as shown in Figure 4-8.
Figure 4-8. Before restructuring, a monolith includes several parts
This system consists of several domain behaviors (identified in the figure as simple
geometric shapes) without much internal organization. In addition, in this scenario,
the desired goal consists of two teams to create two services, one with the hexagon
and square domain, and another with the circle domain, from the existing monolith.
74
|
Chapter 4: Architectural Decomposition
The first step in tactical forking involves cloning the entire monolith, and giving each
team a copy of the entire codebase, as illustrated in Figure 4-9.
Figure 4-9. Step one clones the monolith
Each team receives a copy of the entire codebase, and they start deleting (as illustrated
previously in Figure 4-7) the code they don’t need rather than extract the desirable
code. Developers often find this easier in a tightly coupled codebase because they
don’t have to worry about extracting the large number of dependencies that high cou‐
pling creates. Rather, in the deletion strategy, once functionality has been isolated,
delete any code that doesn’t break anything.
Tactical Forking
|
75
As the pattern continues to progress, teams begin to isolate the target portions, as
shown in Figure 4-10. Then each team continues the gradual elimination of unwan‐
ted code.
Figure 4-10. Teams constantly refactor to remove unwanted code
76
|
Chapter 4: Architectural Decomposition
At the completion of the tactical forking pattern, teams have split the original mono‐
lithic application into two parts, preserving the coarse-grained structure of the behav‐
ior in each part, as illustrated in Figure 4-11.
Figure 4-11. The end state of tactical forking features two services
Now the restructuring is complete, leaving two coarse-grained services as the result.
Trade-Offs
Tactical forking is a viable alternative to a more formal decomposition approach,
most suited to codebases that have little or no internal structure. Like all practices in
architecture, it has its share of trade-offs:
Benefits
• Teams can start working right away with virtually no up-front analysis.
• Developers find it easier to delete code rather than extract it. Extracting code
from a chaotic codebase presents difficulties because of high coupling, whereas
code not needed can be verified by compilation or simple testing.
Tactical Forking
|
77
Shortcomings
• The resulting services will likely still contain a large amount of mostly latent code
left over from the monolith.
• Unless developers undertake additional efforts, the code inside the newly derived
services won’t be better than the chaotic code from the monolith—there’s just less
of it.
• Inconsistencies may occur between the naming of shared code and shared com‐
ponent files, resulting in difficultly identifying common code and keeping it
consistent.
The name of this pattern is apt (as all good pattern names should be)—it provides a
tactical rather than strategic approach for restructuring architectures, allowing teams
to quickly migrate important or critical systems to the next generation (albeit in an
unstructured way).
Sysops Squad Saga: Choosing a Decomposition Approach
Friday, October 29, 10:01
Now that Addison and Austen understood both approaches, they met in the main conference room to analyze the Sysops Squad application using the abstractness and instability metrics to determine which approach would be the most appropriate given their
situation.
“Look at this,” said Addison. “Most of the code lies along the main sequence. There are a
few outliers of course, but I think we can conclude that it’s feasible to break apart this
application. So the next step is to determine which approach to use.”
“I really like the tactical forking approach,” said Austen. “It reminds me of famous sculptors, when
asked how they were able to carve such beautiful works out of solid marble, who replied that they
were merely removing the marble that wasn’t supposed to be there. I feel like the Sysops Squad
application could be my sculpture!”
“Hold on there, Michelangelo,” said Addison. “First sports, and now sculpting? You need to make up
your mind about what you like to spend your nonworking time on. The thing I don’t like about the
tactical forking approach is all the duplicate code and shared functionality within each service. Most
of our problems have to do with maintainability, testability, and overall reliability. Can you imagine
having to apply the same change to several different services at the same time? That would be a
nightmare!”
“But how much shared functionality is there, really?” asked Austen.
“I’m not sure,” said Addison, “but I do know there’s quite a bit of shared code for the infrastructure
stuff like logging and security, and I know a lot of the database calls are shared from the persistence
layer of the application.”
78
|
Chapter 4: Architectural Decomposition
Austen paused and thought about Addison’s argument for a bit. “Maybe you’re right. Since we have
good component boundaries already defined, I’m OK with doing the slower component-based
decomposition approach and giving up my sculpting career. But I’m not giving up sports!”
Addison and Austen came to an agreement that the component decomposition approach would
be the appropriate one for the Sysops Squad application. Addison wrote an ADR for this decision,
outlining the trade-offs and justification for the component-based decomposition approach.
ADR: Migration Using the Component-Based Decomposition Approach
Context
We will be breaking apart the monolithic Sysops Squad application into separately
deployed services. The two approaches we considered for the migration to a distributed
architecture were tactical forking and component-based decomposition.
Decision
We will use the component-based decomposition approach to migrate the existing monolithic Sysops Squad application to a distributed architecture.
The application has well-defined component boundaries, lending itself to the
component-based decomposition approach.
This approach reduces the chance of having to maintain duplicate code within each
service.
With the tactical forking approach, we would have to define the service boundaries up
front to know how many forked applications to create. With the component-based
decomposition approach, the service definitions will naturally emerge through component grouping.
Given the nature of the problems we are facing with the current application with regard to
reliability, availability, scalability, and workflow, using the component-based decomposition approach provides a safer and more controlled incremental migration than the tactical forking approach does.
Consequences
The migration effort will likely take longer with the component-based decomposition
approach than with tactical forking. However, we feel the justifications in the previous section outweigh this trade-off.
This approach allows the developers on the team to work collaboratively to identify
shared functionality, component boundaries, and domain boundaries. Tactical forking
would require us to break apart the team into smaller, separate teams for each forked
application and increase the amount of coordination needed between the smaller teams.
Sysops Squad Saga: Choosing a Decomposition Approach
|
79
CHAPTER 5
Component-Based Decomposition Patterns
Monday, November 1, 11:53
Addison and Austen chose to use the component-based decomposition approach, but
were unsure about the details of each decomposition pattern. They tried to research this
approach, but did not find much on the internet about it. Once again, they met with
Logan in the conference room for advice on what these patterns are all about and how
to use them.
“Listen, Logan,” said Addison, “I want to start out by saying we both really appreciate the
amount of time you have been spending with us to get this migration process started. I
know you’re super busy on your own firefights.”
“No problem,” said Logan. “Us firefighters have to stick together. I’ve been in your shoes before, so I
know what it’s like flying blind on these sort of things. Besides, this is a highly visible migration effort,
and it’s important you both get this thing right the first time. Because there won’t be a second time.”
“Thanks, Logan,” said Austen. “I’ve got a game in about two hours, so we’ll try to make this short. You
talked earlier about component-based decomposition, and we chose that approach, but we aren’t
able to find much about it on the internet.”
“I’m not surprised,” said Logan. “Not much has been written about them yet, but I know a book is
coming out describing these patterns in detail sometime later this year. I first learned about these
decomposition patterns at a conference about four years ago in a session with an experienced software architect. I was really impressed with the iterative and methodical approach to safely move
from a monolithic architecture to a distributed one like service-based architecture and microservices. Since then I’ve been using these patterns with quite a bit of success.”
“Can you show us how these patterns work?” asked Addison.
“Sure,” said Logan. “Let’s take it one pattern at a time.”
81
Component-based decomposition (introduced in Chapter 4) is a highly effective
technique for breaking apart a monolithic application when the codebase is some‐
what structured and grouped by namespaces (or directories). This chapter introduces
a set of patterns, known as component-based decomposition patterns, that describe the
refactoring of monolithic source code to arrive at a set of well-defined components
that can eventually become services. These decomposition patterns significantly ease
the effort of migrating monolithic applications to distributed architectures.
Figure 5-1 shows the road map for the component-based decomposition patterns
described in this chapter and how they are used together to break apart a monolithic
application. Initially, these patterns are used together in sequence when moving a
monolithic application to a distributed one, and then individually as maintenance is
applied to the monolithic application during migration. These decomposition pat‐
terns are summarized as follows:
“Identify and Size Components Pattern” on page 84
Typically the first pattern applied when breaking apart a monolithic application.
This pattern is used to identify, manage, and properly size components.
“Gather Common Domain Components Pattern” on page 94
Used to consolidate common business domain logic that might be duplicated
across the application, reducing the number of potentially duplicate services in
the resulting distributed architecture.
“Flatten Components Pattern” on page 101
Used to collapse or expand domains, subdomains, and components, thus ensur‐
ing that source code files reside only within well-defined components.
“Determine Component Dependencies Pattern” on page 111
Used to identify component dependencies, refine those dependencies, and deter‐
mine the feasibility and overall level of effort for a migration from a monolithic
architecture to a distributed one.
“Create Component Domains Pattern” on page 120
Used to group components into logical domains within the application and to
refactor component namespaces and/or directories to align with a particular
domain.
“Create Domain Services Pattern” on page 126
Used to physically break apart a monolithic architecture by moving logical
domains within the monolithic application to separately deployed domain
services.
82
|
Chapter 5: Component-Based Decomposition Patterns
Figure 5-1. Component-based decomposition pattern flow and usage
Each pattern described in this chapter is divided into three sections. The first section,
“Pattern Description,” describes how the pattern works, why the pattern is important,
and what the outcome is of applying the pattern. Knowing that most systems are
moving targets during a migration, the second section, “Fitness Functions for Gover‐
nance,” describes the automated governance that can be used after applying the pat‐
tern to continually analyze and verify the correctness of the codebase during ongoing
maintenance. The third section uses the real-world Sysops Squad application (see
“Introducing the Sysops Squad Saga” on page 15) to illustrate the use of the pattern
and illustrate the transformations of the application after the pattern has been
applied.
Component-Based Decomposition Patterns
|
83
Architecture Stories
Throughout this chapter, we will be using architecture stories as a way of recording
and describing code refactoring that impacts the structural aspect of the application
for each of the Sysops Squad sagas. Unlike user stories, which describe a feature that
needs to be implemented or changed, an architecture story describes particular code
refactoring that impacts the overall structure of an application and satisfies some sort
of business driver (such as increased scalability, better time-to-market, etc.). For
example, if an architect sees the need to break apart a payment service to support bet‐
ter overall extensibility for adding additional payment types, a new architecture story
would be created and read as follows:
As an architect, I need to decouple the payment service to support better extensibility
and agility when adding additional payment types.
We view architecture stories as separate from technical debt stories. Technical debt
stories usually capture things a developer needs to do in a later iteration to “clean up
the code,” whereas an architecture story captures something that needs to change
quickly to support a particular architectural characteristic or business need.
Identify and Size Components Pattern
The first step in any monolithic migration is to apply the Identify and Size Compo‐
nents pattern. The purpose of this pattern is to identify and catalog the architectural
components (logical building blocks) of the application and then properly size the
components.
Pattern Description
Because services are built from components, it is critical to not only identify the com‐
ponents within an application, but to properly size them as well. This pattern is used
to identify components that are either too big (doing too much) or too small (not
doing enough). Components that are too large relative to other components are gen‐
erally more coupled to other components, are harder to break into separate services,
and lead to a less modular architecture.
Unfortunately, it is difficult to determine the size of a component. The number of
source files, classes, and total lines of code are not good metrics because every pro‐
grammer designs classes, methods, and functions differently. One metric we’ve found
useful for component sizing is calculating the total number of statements within a
given component (the sum of statements within all source files contained within a
namespace or directory). A statement is a single complete action performed in the
source code, usually terminated by a special character (such as a semicolon in lan‐
guages such as Java, C, C++, C#, Go, and JavaScript; or a newline in languages such as
84
|
Chapter 5: Component-Based Decomposition Patterns
F#, Python, and Ruby). While not a perfect metric, at least it’s a good indicator of how
much the component is doing and how complex the component is.
Having a relatively consistent component size within an application is important.
Generally speaking, the size of components in an application should fall between one
to two standard deviations from the average (or mean) component size. In addition,
the percentage of code represented by each component should be somewhat evenly
distributed between application components and not vary significantly.
While many static code analysis tools can show the number of statements within a
source file, many of them don’t accumulate total statement by component. Because of
this, the architect usually must perform manual or automated post-processing to
accumulate total statements by component and then calculate the percentage of code
that component represents.
Regardless of the tools or algorithms used, the important information and metrics to
gather and calculate for this pattern are shown in Table 5-1 and are defined in the
following list.
Table 5-1. Component inventory and component size analysis example
Component name
Component namespace
Percent
Statements
Files
Billing Payment
ss.billing.payment
5
4,312
23
Billing History
ss.billing.history
4
3,209
17
ss.customer.notification
2
1,433
7
Customer Notification
Component name
A descriptive name and identifier of the component that is consistent throughout
application diagrams and documentation. The component name should be clear
enough to be as self-describing as possible. For example, the component Billing
History shown in Table 5-1 is clearly a component that contains source code files
used to manage a customer’s billing history. If the distinct role and responsibility
of the component isn’t immediately identifiable, consider changing the compo‐
nent (and potentially the corresponding namespace) to a more descriptive one.
For example, a component named Ticket Manager leaves too many unanswered
questions about its role and responsibility in the system, and should be renamed
to better describe its role.
Component namespace
The physical (or logical) identification of the component representing where the
source code files implementing that component are grouped and stored. This
identifier is usually denoted through a namespace, package structure (Java), or
directory structure. When a directory structure is used to denote the component,
we usually convert the file separator to a dot (.) and create a corresponding
Identify and Size Components Pattern
|
85
logical namespace. For example, the component namespace for source code files
in the ss/customer/notification directory structure would have the namespace
value ss.customer.notification. Some languages require that the namespace
match the directory structure (such as Java with a package), whereas other lan‐
guages (such as C# with a namespace) do not enforce this constraint. Whatever
namespace identifier is used, make sure the type of identifier is consistent across
all of the components in the application.
Percent
The relative size of the component based on its percentage of the overall source
code containing that component. The percent metric is helpful in identifying
components that appear too large or too small in the overall application. This
metric is calculated by taking the total number of statements within the source
code files representing that component and dividing that number by the total
number of statements in the entire codebase of the application. For example, the
percent value of 5 for the ss.billing.payment component in Table 5-1 means
that this component constitutes 5% of the overall codebase.
Statements
The sum of the total number of source code statements in all source files con‐
tained within that component. This metric is useful for determining not only the
relative size of the components within an application, but also for determining
the overall complexity of the component. For example, a seemingly simple singlepurpose component named Customer Wishlist might have a total of 12,000 state‐
ments, indicating that the processing of wish list items is perhaps more complex
than it looks. This metric is also necessary for calculating the percent metric pre‐
viously described.
Files
The total number of source code files (such as classes, interfaces, types, and so
on) that are contained within the component. While this metric has little to do
with the size of a component, it does provide additional information about the
component from a class structure standpoint. For example, a component with
18,409 statements and only 2 files is a good candidate for refactoring into smaller,
more contextual classes.
When resizing a large component, we recommend using a functional decomposition
approach or a domain-driven approach to identify subdomains that might exist
within the large component. For example, assume the Sysops Squad application has a
Trouble Ticket component containing 22% of the codebase that is responsible for
ticket creation, assignment, routing, and completion. In this case, it might make sense
to break the single Trouble Ticket component into four separate components (Ticket
Creation, Ticket Assignment, Ticket Routing, and Ticket Completion), reducing the
percentage of code each component represents, therefore creating a more modular
86
| Chapter 5: Component-Based Decomposition Patterns
application. If no clear subdomains exist within a large component, then leave the
component as is.
Fitness Functions for Governance
Once this decomposition pattern has been applied and components have been identi‐
fied and sized correctly, it’s important to apply some sort of automated governance to
identify new components and to ensure components don’t get too large during nor‐
mal application maintenance and create unwanted or unintended dependencies.
Automated holistic fitness functions can be triggered during deployment to alert the
architect if specified constraints are exceeded (such as the percent metric discussed
previously or use of standard deviations to identify outliers).
Fitness functions can be implemented through custom-written code or through the
use of open source or COTS tools as part of a CI/CD pipeline. Some of the automated
fitness functions that can be used to help govern this decomposition pattern are as
follows.
Fitness function: Maintain component inventory
This automated holistic fitness function, usually triggered on deployment through a
CI/CD pipeline, helps keep the inventory of components current. It’s used to alert an
architect of components that might have been added or removed by the development
team. Identifying new or removed components is not only critical for this pattern, but
for the other decomposition patterns as well. Example 5-1 shows the pseudocode and
algorithm for one possible implementation of this fitness function.
Example 5-1. Pseudocode for maintaining component inventory
# Get prior component namespaces that are stored in a datastore
LIST prior_list = read_from_datastore()
# Walk the directory structure, creating namespaces for each complete path
LIST current_list = identify_components(root_directory)
# Send an alert if new or removed components are identified
LIST added_list = find_added(current_list, prior_list)
LIST removed_list = find_removed(current_list, prior_list)
IF added_list NOT EMPTY {
add_to_datastore(added_list)
send_alert(added_list)
}
IF removed_list NOT EMPTY {
remove_from_datastore(removed_list)
send_alert(removed_list)
}
Identify and Size Components Pattern
|
87
Fitness function: No component shall exceed <some percent> of the overall codebase
This automated holistic fitness function, usually triggered on deployment through a
CI/CD pipeline, identifies components that exceed a given threshold in terms of the
percentage of overall source code represented by that component, and alerts the
architect if any component exceeds that threshold. As mentioned earlier in this chap‐
ter, the threshold percentage value will vary depending on the size of the application,
but should be set so as to identify significant outliers. For example, for a relatively
small application with only 10 components, setting the percentage threshold to some‐
thing like 30% would sufficiently identify a component that is too large, whereas for a
large application with 50 components, a threshold of 10% would be more appropri‐
ate. Example 5-2 shows the pseudocode and algorithm for one possible implementa‐
tion of this fitness function.
Example 5-2. Pseudocode for maintaining component size based on percent of code
# Walk the directory structure, creating namespaces for each complete path
LIST component_list = identify_components(root_directory)
# Walk through all of the source code to accumulate total statements
total_statements = accumulate_statements(root_directory)
# Walk through the source code for each component, accumulating statements
# and calculating the percentage of code each component represents. Send
# an alert if greater than 10%
FOREACH component IN component_list {
component_statements = accumulate_statements(component)
percent = component_statements / total_statements
IF percent > .10 {
send_alert(component, percent)
}
}
Fitness function: No component shall exceed <some number of standard deviations> from the
mean component size
This automated holistic fitness function, usually triggered on deployment through a
CI/CD pipeline, identifies components that exceed a given threshold in terms of the
number of standard deviations from the mean of all component sizes (based on the
total number of statements in the component), and alerts the architect if any compo‐
nent exceeds that threshold.
Standard deviation is a useful means of determining outliers in terms of component
size. Standard deviation is calculated as follows:
s=
88
|
1
N
x −x
N − 1 ∑i = 1 i
2
Chapter 5: Component-Based Decomposition Patterns
where N is the number of observed values, xi is the observed values, and x is the mean
of the observed values. The mean of observed values (x) is calculated as follows:
x=
1 N
xi
N i∑
=1
The standard deviation can then be used along with the difference from the mean to
determine the number of standard deviations the component size is from the mean.
Example 5-3 shows the pseudocode for this fitness function, using three standard
deviations from the mean as a threshold.
Example 5-3. Pseudocode for maintaining component size based on number of standard
deviations
# Walk the directory structure, creating namespaces for each complete path
LIST component_list = identify_components(root_directory)
# Walk through all of the source code to accumulate total statements and number
# of statements per component
SET total_statements TO 0
MAP component_size_map
FOREACH component IN component_list {
num_statements = accumulate_statements(component)
ADD num_statements TO total_statements
ADD component,num_statements TO component_size_map
}
# Calculate the standard deviation
SET square_diff_sum TO 0
num_components = get_num_entries(component_list)
mean = total_statements / num_components
FOREACH component,size IN component_size_map {
diff = size - mean
ADD square(diff) TO square_diff_sum
}
std_dev = square_root(square_diff_sum / (num_components - 1))
# For each component calculate the number of standard deviations from the
# mean. Send an alert if greater than 3
FOREACH component,size IN component_size_map {
diff_from_mean = absolute_value(size - mean);
num_std_devs = diff_from_mean / std_dev
IF num_std_devs > 3 {
send_alert(component, num_std_devs)
}
}
Identify and Size Components Pattern
|
89
Sysops Squad Saga: Sizing Components
Tuesday, November 2, 09:12
After the discussion with Logan (the lead architect) about component-based decomposition patterns, Addison decided to apply the Identify and Size Components pattern to
identify all of the components in the Sysops Squad ticketing application and calculate
the size of each component based on the total number of statements in each
component.
Addison gathered all the necessary component information and put this information
into Table 5-2, calculating the percentage of code for each component based on the
total number of statements in the entire application (in this case, 82,931 statements).
Table 5-2. Component size analysis for the Sysops Squad application
Component name
Component namespace
Percent
Statements
Files
Login
ss.login
2
1865
3
Billing Payment
ss.billing.payment
5
4,312
23
Billing History
ss.billing.history
4
3,209
17
Customer Notification
ss.customer.notification
2
1,433
7
Customer Profile
ss.customer.profile
5
4,012
16
Expert Profile
ss.expert.profile
6
5,099
32
KB Maint
ss.kb.maintenance
2
1,701
14
KB Search
ss.kb.search
3
2,871
4
Reporting
ss.reporting
33
27,765
162
Ticket
ss.ticket
8
7,009
45
Ticket Assign
ss.ticket.assign
9
7,845
14
Ticket Notify
ss.ticket.notify
2
1,765
3
Ticket Route
ss.ticket.route
2
1,468
4
Support Contract
ss.supportcontract
5
4,104
24
Survey
ss.survey
3
2,204
5
Survey Notify
ss.survey.notify
2
1,299
3
Survey Templates
ss.survey.templates
2
1,672
7
User Maintenance
ss.users
4
3,298
12
90
|
Chapter 5: Component-Based Decomposition Patterns
Addison noticed that most of the components listed in Table 5-2 are about the same size, with the
exception of the Reporting component (ss.reporting) which consisted of 33% of the codebase.
Since the Reporting component was significantly larger than the other components (illustrated in
Figure 5-2), Addison chose to break this component apart to reduce its overall size.
Figure 5-2. The Reporting component is too big and should be broken apart
After doing some analysis, Addison found that the reporting component contained source code
that implemented three categories of reports:
• Ticketing reports (ticket demographics reports, tickets per day/week/month reports, ticket resolution time reports, and so on)
• Expert reports (expert utilization reports, expert distribution reports, and so on)
Identify and Size Components Pattern
|
91
• Financial reports (repair cost reports, expert cost reports, profit reports, and so on)
Addison also identified common (shared) code that all reporting categories used, such as common
utilities, calculators, shared data queries, report distribution, and shared data formatters. Addison created an architecture story (see “Architecture Stories” on page 84) for this refactoring and explained it
to the development team. Sydney, one of the Sysops Squad developers assigned the architecture
story, refactored the code to break apart the single Reporting component into four separate components—a Reporting Shared component containing the common code and three other components
(Ticket Reports, Expert Reports, and Financial Reports), each representing a functional reporting
area, as illustrated in Figure 5-3.
Figure 5-3. The large Reporting component broken into smaller reporting components
92
| Chapter 5: Component-Based Decomposition Patterns
After Sydney committed the changes, Addison reanalyzed the code and verified that all of the components were now fairly equally distributed in size. Addison recorded the results of applying this
decomposition pattern in Table 5-3.
Table 5-3. Component size after applying the Identify and Size Components pattern
Component name
Component namespace
Percent
Statements
Files
Login
ss.login
2
1865
3
Billing Payment
ss.billing.payment
5
4,312
23
Billing History
ss.billing.history
4
3,209
17
Customer Notification
ss.customer.notification
2
1,433
7
Customer Profile
ss.customer.profile
5
4,012
16
Expert Profile
ss.expert.profile
6
5,099
32
KB Maint
ss.kb.maintenance
2
1,701
14
KB Search
ss.kb.search
3
2,871
4
Reporting Shared
ss.reporting.shared
7
5,309
20
Ticket Reports
ss.reporting.tickets
8
6,955
58
Expert Reports
ss.reporting.experts
9
7,734
48
Financial Reports
ss.reporting.financial
9
7,767
36
Ticket
ss.ticket
8
7,009
45
Ticket Assign
ss.ticket.assign
9
7,845
14
Ticket Notify
ss.ticket.notify
2
1,765
3
Ticket Route
ss.ticket.route
2
1,468
4
Support Contract
ss.supportcontract
5
4,104
24
Survey
ss.survey
3
2,204
5
Survey Notify
ss.survey.notify
2
1,299
3
Survey Templates
ss.survey.templates
2
1,672
7
User Maintenance
ss.users
4
3,298
12
Notice in the preceding Sysops Squad Saga that Reporting no longer exists as a com‐
ponent in Table 5-3 or Figure 5-3. Although the namespace still exists (ss.report
ing), it is no longer considered a component, but rather a subdomain. The refactored
components listed in Table 5-3 will be used when applying the next decomposition
pattern, Gather Common Domain Components.
Identify and Size Components Pattern
|
93
Gather Common Domain Components Pattern
When moving from a monolithic architecture to a distributed one, it is often benefi‐
cial to identify and consolidate common domain functionality to make common
services easier to identify and create. The Gather Common Domain Components pat‐
tern is used to identify and collect common domain logic and centralize it into a sin‐
gle component.
Pattern Description
Shared domain functionality is distinguished from shared infrastructure functionality
in that domain functionality is part of the business processing logic of an application
(such as notification, data formatting, and data validation) and is common to only
some processes, whereas infrastructure functionality is operational in nature (such as
logging, metrics gathering, and security) and is common to all processes.
Consolidating common domain functionality helps eliminate duplicate services when
breaking apart a monolithic system. Often there are only very subtle differences
among common domain functionality that is duplicated throughout the application,
and these differences can be easily resolved within a single common service (or
shared library).
Finding common domain functionality is mostly a manual process, but some auto‐
mation can be used to assist in this effort (see “Fitness Functions for Governance” on
page 95). One hint that common domain processing exists in the application is the
use of shared classes across components or a common inheritance structure used by
multiple components. Take, for example, a class file named SMTPConnection in a
large codebase that is used by five classes, all contained within different namespaces
(components). This scenario is a good indication that common email notification
functionality is spread throughout the application and might be a good candidate for
consolidation.
Another way of identifying common domain functionality is through the name of a
logical component or its corresponding namespace. Consider the following compo‐
nents (represented as namespaces) in a large codebase:
• Ticket Auditing (penultimate.ss.ticket.audit)
• Billing Auditing (penultimate.ss.billing.audit)
• Survey Auditing (penultimate.ss.survey.audit)
94
|
Chapter 5: Component-Based Decomposition Patterns
Notice how each of these components (Ticket Auditing, Billing Auditing, and Survey
Auditing) all have the same thing in common—writing the action performed and the
user requesting the action to an audit table. While the context may be different, the
final outcome is the same—inserting a row in an audit table. This common domain
functionality can be consolidated into a new component called penulti
mate.ss.shared.audit, resulting in less duplication of code and also fewer services
in the resulting distributed architecture.
Not all common domain functionality necessarily becomes a shared service. Alterna‐
tively, common code could be gathered into a shared library that is bound to the code
during compile time. The pros and cons of using a shared service rather than a
shared library are discussed in detail in Chapter 8.
Fitness Functions for Governance
Automating the governance of shared domain functionality is rather difficult because
of the subjectiveness of identifying shared functionality and classifying it as domain
functionality versus infrastructure functionality. For the most part, the fitness func‐
tions used to govern this pattern are therefore somewhat manual. That said, there are
some ways to automate the governance to assist in the manual interpretation of com‐
mon domain functionality. The following fitness functions can assist in finding com‐
mon domain functionality.
Fitness function: Find common names in leaf nodes of component namespace
This automated holistic fitness function can be triggered on deployment through a
CI/CD pipeline to locate common names within the namespace of a component.
When a common ending namespace node name is found between two or more com‐
ponents, the architect is alerted and can analyze the functionality to determine if it is
common domain logic. So that the same alert isn’t continuously sent as a “false posi‐
tive,” an exclusion file can be used to store those namespaces that have common end‐
ing node names but are not deemed common domain logic (such as multiple
namespaces ending in .calculate or .validate). Example 5-4 shows the pseudo‐
code for this fitness function.
Example 5-4. Pseudocode for finding common namespace leaf node names
# Walk the directory structure, creating namespaces for each complete path
LIST component_list = identify_components(root_directory)
# Locate possible duplicate component node names that are not in the exclusion
# list stored in a datastore
LIST excluded_leaf_node_list = read_datastore()
LIST leaf_node_list
LIST common_component_list
Gather Common Domain Components Pattern
|
95
FOREACH component IN component_list {
leaf_name = get_last_node(component)
IF leaf_name IN leaf_node_list AND
leaf_name NOT IN excluded_leaf_node_list {
ADD component TO common_component_list
} ELSE {
ADD leaf_name TO leaf_node_list
}
}
# Send an alert if any possible common components were found
IF common_component_list NOT EMPTY {
send_alert(common_component_list)
}
Fitness function: Find common code across components
This automated holistic fitness function can be triggered on deployment through a
CI/CD pipeline to locate common classes used between namespaces. While not
always accurate, it does help in alerting an architect of possible duplicate domain
functionality. Like the previous fitness function, an exclusion file is used to reduce the
number of “false positives” for known common code that is not considered duplicate
domain logic. Example 5-5 shows the pseudocode for thisfitness function.
Example 5-5. Pseudocode for finding common source files between components
# Walk the directory structure, creating namespaces for each complete path and a list
# of source file names for each component
LIST component_list = identify_components(root_directory)
LIST source_file_list = get_source_files(root_directory)
MAP component_source_file_map
FOREACH component IN component_list {
LIST component_source_file_list = get_source_files(component)
ADD component, component_source_file_list TO component_source_file_map
}
# Locate possible common source file usage across components that are not in
# the exclusion list stored in a datastore
LIST excluded_source_file_list = read_datastore()
LIST common_source_file_list
FOREACH source_file IN source_file_list {
SET count TO 0
FOREACH component,component_source_file_list IN component_source_file_map {
IF source_file IN component_source_file_list {
ADD 1 TO count
}
}
IF count > 1 AND source_file NOT IN excluded_source_file_list {
ADD source_file TO common_source_file_list
}
96
|
Chapter 5: Component-Based Decomposition Patterns
}
# Send an alert if any source files are used in multiple components
IF common_source_file_list NOT EMPTY {
send_alert(common_source_file_list)
}
Sysops Squad Saga: Gathering Common Components
Friday, November 5, 10:34
Having identified and sized the components in the Sysops Squad application, Addison
applied the Gather Common Domain Components pattern to see if any common functionality existed between components. From the list of components in Table 5-3, Addison noticed there were three components all related to notifying a Sysops Squad
customer, and listed these in Table 5-4.
Table 5-4. Sysops Squad components with common domain functionality
Component
Namespace
Responsibility
Customer Notification
ss.customer.notification
General notification
Ticket Notify
ss.ticket.notify
Notify that expert is en route
Survey Notify
ss.survey.notify
Send survey email
While each of these notification components had a different context for notifying a customer, Addison realized they all have one thing in common—they all sent information to a customer. Figure 5-4
illustrates these common notification components within the Sysops Squad application.
Noticing that the source code contained in these components was also very similar, Addison consulted with Austen (the other Sysops Squad architect). Austen liked the idea of a single notification
component, but was concerned about impacting the overall level of coupling between components. Addison agreed that this might be an issue and investigated this trade-off further.
Gather Common Domain Components Pattern
|
97
Figure 5-4. Notification functionality is duplicated throughout the application
Addison analyzed the incoming (afferent) coupling level for the existing Sysops Squad notification
components and came up with the resulting coupling metrics listed in Table 5-5, with “CA” representing the number of other components requiring that component (afferent coupling).
98
| Chapter 5: Component-Based Decomposition Patterns
Table 5-5. Sysops Squad coupling analysis before component consolidation
Component
CA
Used by
Customer Notification
2
Billing Payment, Support Contract
Ticket Notify
2
Ticket, Ticket Route
Survey Notify
1
Survey
Addison then found that if the customer notification functionality was consolidated into a single
component, the coupling level for the resulting single component increased to an incoming coupling level of 5, as shown in Table 5-6.
Table 5-6. Sysops Squad coupling analysis after component consolidation
Component
CA
Used by
Notification
5
Billing Payment, Support Contract, Ticket, Ticket Route, Survey
Addison brought these findings to Austen, and they discussed the results. What they found is that,
while the new consolidated component had a fairly high level of incoming coupling, it didn’t affect
the overall afferent (incoming) coupling level for notifying a customer. In other words, the three separate components had a total incoming coupling level of 5, but so did the single consolidated
component.
Addison and Austen both realized how important it was to analyze the coupling level after consolidating common domain functionality. In some cases, combining common domain functionality into
a single consolidated component increased the incoming coupling level of that component, thus
resulting in too many dependencies on a single shared component within the application. However,
in this case both Addison and Austen were comfortable with the coupling analysis, and agreed to
consolidate the notification functionality to reduce the duplication of both code and functionality.
Addison wrote an architecture story to combine all of the notification functionality into a single
namespace representing a common Notification component. Sydney, assigned to the architecture
story, refactored the source code, creating a single component for customer notification, as illustrated in Figure 5-5.
Gather Common Domain Components Pattern
|
99
Figure 5-5. Notification functionality is consolidated into a new single component called
Notification
Table 5-7 shows the resulting components after Sydney implemented the architecture story Addison created. Notice that the Customer Notification component (ss.customer.notification), Ticket
Notify component (ss.ticket.notify), and Survey Notify components (ss.survey.notify) were
removed, and the source code moved to the new consolidated Notification component
( ss.notification).
100
|
Chapter 5: Component-Based Decomposition Patterns
Table 5-7. Sysops Squad components after applying the Gather Common Domain
Components pattern
Component
Namespace
Responsibility
Login
ss.login
User and customer login
Billing Payment
ss.billing.payment
Customer monthly billing
Billing History
ss.billing.history
Payment history
Customer Profile
ss.customer.profile
Maintain customer profile
Expert Profile
ss.expert.profile
Maintain expert profile
KB Maint
ss.kb.maintenance
Maintain & view knowledge base
KB Search
ss.kb.search
Search knowledge base
Notification
ss.notification
All customer notification
Reporting Shared
ss.reporting.shared
Shared functionality
Ticket Reports
ss.reporting.tickets
Create ticketing reports
Expert Reports
ss.reporting.experts
Create expert reports
Financial Reports
ss.reporting.financial
Create financial reports
Ticket
ss.ticket
Ticket creation & maintenance
Ticket Assign
ss.ticket.assign
Assign expert to ticket
Ticket Route
ss.ticket.route
Send ticket to expert
Support Contract
ss.supportcontract
Support contract maintenance
Survey
ss.survey
Send and receive surveys
Survey Templates
ss.survey.templates
Maintain survey templates
User Maintenance
ss.users
Maintain internal users
Flatten Components Pattern
As mentioned previously, components—the building blocks of an application—are
usually identified through namespaces, package structures, or directory structures
and are implemented through class files (or source code files) contained within these
structures. However, when components are built on top of other components, which
are in turn built on top of other components, they start to lose their identity and stop
becoming components as per our definition. The Flatten Components pattern is used
to ensure that components are not built on top of one another, but rather flattened
and represented as leaf nodes in a directory structure or namespace.
Flatten Components Pattern
|
101
Pattern Description
When the namespace representing a particular component gets extended (in other
words, another node is added to the namespace or directory structure), the prior
namespace or directory no longer represents a component, but rather a subdomain.
To illustrate this point, consider the customer survey functionality within the Sysops
Squad application represented by two components: Survey (ss.survey) and Survey
Templates (ss.survey.templates). Notice in Table 5-8 how the ss.survey name‐
space, which contains five class files used to manage and collect the surveys, is exten‐
ded with the ss.survey.templates namespace to include seven classes representing
each survey type send out to customers.
Table 5-8. The Survey component contains orphaned classes and should be flattened
Component name
Component namespace
Files
→ Survey
ss.survey
5
Survey Templates
ss.survey.templates
7
While this structure might seem to make sense from a developer’s standpoint in order
to keep the template code separate from survey processing, it does create some prob‐
lems because Survey Templates, as a component, would be considered part of the Sur‐
vey component. One might be tempted to consider Survey Templates as a
subcomponent of Survey, but then issues arise when trying to form services from these
components—should both components reside in a single service called Survey, or
should the Survey Templates be a separate service from the Survey service?
We’ve resolved this dilemma by defining a component as the last node (or leaf node)
of the namespace or directory structure. With this definition, ss.survey.templates
is a component, whereas ss.survey would be considered a subdomain, not a compo‐
nent. We further define namespaces such as ss.survey as root namespaces because
they are extended with other namespace nodes (in this case, .templates).
Notice how the ss.survey root namespace in Table 5-8 contains five class files. We
call these class files orphaned classes because they do not belong to any definable com‐
ponent. Recall that a component is identified by a leaf node namespace containing
source code. Because the ss.survey namespace was extended to include .templates,
ss.survey is no longer considered a component and therefore should not contain
any class files.
102
|
Chapter 5: Component-Based Decomposition Patterns
The following terms and corresponding definitions are important for understanding
and applying the Flatten Components decomposition pattern:
Component
A collection of classes grouped within a leaf node namespace that performs some
sort of specific functionality in the application (such as payment processing or
customer survey functionality).
Root namespace
A namespace node that has been extended by another namespace node. For
example, given the namespaces ss.survey and ss.survey.templates, ss.sur
vey would be considered a root namespace because it is extended by .templates.
Root namespaces are also sometimes referred to as subdomains.
Orphaned classes
Classes contained within a root namespace, and hence have no definable compo‐
nent associated with them.
These definitions are illustrated in Figure 5-6, where the box with a C represents
source code contained within that namespace. This diagram (and all others like it) are
purposely drawn from the bottom up to emphasize the notion of hills in the applica‐
tion, as well as emphasize the notion of namespaces building upon each other.
Figure 5-6. Components, root namespaces, and orphaned classes (C box denotes source
code)
Notice that since both ss.survey and ss.ticket are extended through other name‐
space nodes, those namespaces are considered root namespaces, and the classes con‐
tained in those root namespaces are hence orphaned classes (belonging to no defined
Flatten Components Pattern
|
103
component). Thus, the only components denoted in Figure 5-6 are ss.survey.tem
plates, ss.login, ss.ticket.assign, and ss.ticket.route.
The Flatten Components decomposition pattern is used to move orphaned classes to
create well-defined components that exist only as leaf nodes of a directory or name‐
space, creating well-defined subdomains (root namespaces) in the process. We refer
to the flattening of components as the breaking down (or building up) of namespaces
within an application to remove orphaned classes. For example, one way of flattening
the ss.survey root namespace in Figure 5-6 and remove orphaned classes is to move
the source code contained in the ss.survey.templates namespace down to the
ss.survey namespace, thereby making ss.survey a single component (.survey is
now the leaf node of that namespace). This flattening option is illustrated in Figure
5-7.
Figure 5-7. Survey is flattened by moving the survey template code into the .survey
namespace
Alternatively, flattening could also be applied by taking the source code in ss.survey
and applying functional decomposition or domain-driven design to identify separate
functional areas within the root namespace, thus forming components from those
functional areas. For example, suppose the functionality within the ss.survey name‐
space creates and sends a survey to a customer, and then processes a completed sur‐
vey received from the customer. Two components could be created from the
ss.survey namespace: ss.survey.create, which creates and sends the survey, and
ss.survey.process, which processes a survey received from a customer. This form
of flattening is illustrated in Figure 5-8.
104
|
Chapter 5: Component-Based Decomposition Patterns
Figure 5-8. Survey is flattened by moving the orphaned classes to new leaf nodes
( components)
Regardless of the direction of flattening, make sure source code
files reside only in leaf node namespaces or directories so that
source code can always be identified within a specific component.
Another common scenario where orphaned source code might reside in a root name‐
space is when code is shared by other components within that namespace. Consider
the example in Figure 5-9 where customer survey functionality resides in three com‐
ponents (ss.survey.templates, ss.survey.create, and ss.survey.process), but
common code (such as interfaces, abstract classes, common utilities) resides in the
root namespace ss.survey.
Figure 5-9. Shared code in .survey is considered orphaned classes and should be moved
Flatten Components Pattern
|
105
The shared classes in ss.survey would still be considered orphaned classes, even
though they represent shared code. Applying the Flatten Components pattern would
move those shared orphaned classes to a new component called ss.survey.shared,
therefore removing all orphaned classes from the ss.survey subdomain, as illustra‐
ted in Figure 5-10.
Figure 5-10. Shared survey code is moved into its own component
Our advice when moving shared code to a separate component (leaf node name‐
space) is to pick a word that is not used in any existing codebase in the domain, such
as .sharedcode, .commoncode, or some such unique name. This allows the architect
to generate metrics based on the number of shared components in the codebase, as
well as the percentage of source code that is shared in the application. This is a good
indicator as to the feasibility of breaking up the monolithic application. For example,
if the sum of all the statements in all namespaces ending with .sharedcode consti‐
tutes 45% of the overall source code, chances are moving to a distributed architecture
will result in too many shared libraries and end up becoming a nightmare to maintain
because of shared library dependencies.
Another good metric involving the analysis of shared code is the number of compo‐
nents ending in .sharedcode (or whatever common shared namespace node is used).
This metric gives the architect insight into how many shared libraries (JAR, DLL, and
so on) or shared services will result from breaking up the monolithic application.
106
|
Chapter 5: Component-Based Decomposition Patterns
Fitness Functions for Governance
Applying the Flatten Components decomposition pattern involves a fair amount of
subjectivity. For example, should code from leaf nodes be consolidated into the root
namespace, or should code in a root namespace be moved into leaf nodes? That said,
the following fitness function can assist in automating the governance of keeping
components flat (only in leaf nodes).
Fitness function: No source code should reside in a root namespace
This automated holistic fitness function can be triggered on deployment through a
CI/CD pipeline to locate orphaned classes—classes that reside in a root namespace.
Use of this fitness function helps keep components flat when undergoing a mono‐
lithic migration, especially when performing ongoing maintenance to the monolithic
application during the migration effort. Example 5-6 shows the pseudocode that
alerts an architect when orphaned classes appear anywhere in the codebase.
Example 5-6. Pseudocode for finding code in root namespaces
# Walk the directory structure, creating namespaces for each complete path
LIST component_list = identify_components(root_directory)
# Send an alert if a non-leaf node in any component contains source files
FOREACH component IN component_list {
LIST component_node_list = get_nodes(component)
FOREACH node IN component_node_list {
IF contains_code(node) AND NOT last_node(component_node_list) {
send_alert(component)
}
}
}
Sysops Squad Saga: Flattening Components
Wednesday, November 10, 11:10
After applying the “Gather Common Domain Components Pattern” on page 94, Addison
analyzed the results in Table 5-7 and observed that the Survey and Ticket components
contained orphaned classes. Addison highlighted these components in Table 5-9 and in
Figure 5-11.
Flatten Components Pattern
|
107
Figure 5-11. The Survey and Ticket components contain orphaned classes and should be
flattened
Table 5-9. Sysops Squad Ticket and Survey components should be flattened
Component name
Component namespace
Statements
Files
Ticket
ss.ticket
7,009
45
Ticket Assign
ss.ticket.assign
7,845
14
Ticket Route
ss.ticket.route
1,468
4
Survey
ss.survey
2,204
5
Survey Templates
ss.survey.templates
1,672
7
108
|
Chapter 5: Component-Based Decomposition Patterns
Addison decided to address the ticketing components first. Knowing that flattening components
meant getting rid of source code in nonleaf nodes, Addison had two choices: consolidate the code
contained in the ticket assignment and ticket routing components into the ss.ticket component,
or break up the 45 classes in the ss.ticket component into separate components, thus making
ss.ticket a subdomain. Addison discussed these options with Sydney (one of the Sysops Squad
developers), and based on the complexity and frequent changes in the ticket assignment functionality, decided to keep those components separate and move the orphaned code from the
ss.ticket root namespace into other namespaces, thus forming new components.
With help from Sydney, Addison found that the 45 orphaned classes contained in the ss.ticket
namespace implemented the following ticketing functionality:
• Ticket creation and maintenance (creating a ticket, updating a ticket, canceling a ticket, etc.)
• Ticket completion logic
• Shared code common to most of the ticketing functionality
Since ticket assignment and ticket routing functionality were already in their own components
(ss.ticket.assign and ss.ticket.route, respectively), Addison created an architecture story to
move the source code contained in the ss.ticket namespace to three new components, as shown
in Table 5-10.
Table 5-10. The prior Sysops Squad Ticket component broken into three new components
Component
Namespace
Responsibility
Ticket Shared
ss.ticket.shared
Common code and utilities
Ticket Maintenance
ss.ticket.maintenance
Add and maintain tickets
Ticket Completion
ss.ticket.completion
Complete ticket and initiate survey
Ticket Assign
ss.ticket.assign
Assign expert to ticket
Ticket Route
ss.ticket.route
Send ticket to expert
Addison then considered the survey functionality. Working with Sydney, Addison found that the survey functionality rarely changed and was not overly complicated. Sydney talked with Skyler, the
Sysops Squad developer who originally created the ss.survey.templates namespace, and found
there was no compelling reason to separate the survey templates into their own namespace (“It just
seemed like a good idea at the time,” said Skyler). With this information, Addison created an architecture story to move the seven class files from ss.survey.templates into the ss.survey namespace
and removed the ss.survey.template component, as shown in Table 5-11.
Flatten Components Pattern
|
109
Table 5-11. The prior Sysops Squad Survey components flattened into a single component
Component
Namespace
Responsibility
Survey
ss.survey
Send and seceive surveys
After applying the Flatten Components pattern (illustrated in Figure 5-12), Addison observed that
there were no “hills” (component upon component) or orphaned classes and that all of the components were contained only in the leaf nodes of the corresponding namespace.
Figure 5-12. The Survey component was flattened into a single component, whereas the
Ticket component was raised up and flattened, creating a Ticket subdomain
Addison recorded the results of the refactoring efforts thus far in applying these decomposition patterns and listed them in Table 5-12.
110
| Chapter 5: Component-Based Decomposition Patterns
Table 5-12. Sysops Squad components after applying the Flatten Components pattern
Component
Namespace
Login
ss.login
Billing Payment
ss.billing.payment
Billing History
ss.billing.history
Customer Profile
ss.customer.profile
Expert Profile
ss.expert.profile
KB Maint
ss.kb.maintenance
KB Search
ss.kb.search
Notification
ss.notification
Reporting Shared
ss.reporting.shared
Ticket Reports
ss.reporting.tickets
Expert Reports
ss.reporting.experts
Financial Reports
ss.reporting.financial
Ticket Shared
ss.ticket.shared
Ticket Maintenance
ss.ticket.maintenance
Ticket Completion
ss.ticket.completion
Ticket Assign
ss.ticket.assign
Ticket Route
ss.ticket.route
Support Contract
ss.supportcontract
Survey
ss.survey
User Maintenance
ss.users
Determine Component Dependencies Pattern
Three of the most common questions asked when considering a migration from a
monolithic application to a distributed architecture are as follows:
1. Is it feasible to break apart the existing monolithic application?
2. What is the rough overall level of effort for this migration?
3. Is this going to require a rewrite of the code or a refactoring of the code?
Determine Component Dependencies Pattern
|
111
One of your authors was engaged several years ago in a large migration effort to move
a complex monolithic application to microservices. On the first day of the project, the
CIO wanted to know only one thing—was this migration effort a golfball, basketball,
or an airliner? Your author was curious about the sizing comparisons, but the CIO
insisted that the answer to this simple question shouldn’t be that difficult given that
kind of coarse-grained sizing. As it turned out, applying the Determine Component
Dependencies pattern quickly and easily answered this question for the CIO—the
effort was unfortunately an airliner, but only a small Embraer 190 migration rather
than a large Boeing 787 Dreamliner migration.
Pattern Description
The purpose of the Determine Component Dependencies pattern is to analyze the
incoming and outgoing dependencies (coupling) between components to determine
what the resulting service dependency graph might look like after breaking up the
monolithic application. While there are many factors in determining the right level of
granularity for a service (see Chapter 7), each component in the monolithic applica‐
tion is potentially a service candidate (depending on the target distributed architec‐
ture style). For this reason, it is critical to understand the interactions and
dependencies between components.
It’s important to note that this pattern is about component dependencies, not individ‐
ual class dependencies within a component. A component dependency is formed when
a class from one component (namespace) interacts with a class from another compo‐
nent (namespace). For example, suppose the CustomerSurvey class in the ss.survey
component invokes a method in the CustomerNotification class in the ss.notifica
tion component to send out the customer survey, as illustrated in the pseudocode in
Example 5-7.
Example 5-7. Pseudocode showing a dependency between the Survey and Notification
components
namespace ss.survey
class CustomerSurvey {
function createSurvey {
...
}
function sendSurvey {
...
ss.notification.CustomerNotification.send(customer_id, survey)
}
}
112
|
Chapter 5: Component-Based Decomposition Patterns
Notice the dependency between the Survey and Notification components, because the
CustomerNotification class used by the CustomerSurvey class resides outside the
ss.survey namespace. Specifically, the Survey component would have an efferent (or
outgoing) dependency on the Notification component, and the Notification compo‐
nent would have an afferent (or incoming) dependency on the Survey component.
Note that the classes within a particular component may be a highly coupled mess of
numerous dependencies, but that doesn’t matter when applying this pattern—what
matters is only those dependencies between components.
Several tools are available that can assist in applying this pattern and visualizing com‐
ponent dependencies. In addition, many modern IDEs have plug-ins that will pro‐
duce dependency diagrams of the components, or namespaces, within a particular
codebase. These visualizations can be useful in answering the three key questions
posed at the start of this section.
For example, consider the dependency diagram shown in Figure 5-13, where the
boxes represent components (not classes), and the lines represent coupling points
between the components. Notice there is only a single dependency between the com‐
ponents in this diagram, making this application a good candidate for breaking apart
since the components are functionally independent from one another.
Figure 5-13. A monolithic application with minimal component dependencies takes less
effort to break apart (golf ball sizing)
Determine Component Dependencies Pattern
|
113
With a dependency diagram like Figure 5-13, the answers to the three key questions
are as follows:
1. Is it feasible to break apart the existing monolithic application? Yes
2. What is the rough overall level of effort for this migration? A golf ball (relatively
straightforward)
3. Is this going to be a rewrite of the code or a refactoring of the code? Refactoring
(moving existing code into separately deployed services)
Now look at the dependency diagram shown in Figure 5-14. Unfortunately, this dia‐
gram is typical of the dependencies between components in most business applica‐
tions. Notice in particular how the lefthand side of this diagram has the highest level
of coupling, whereas the righthand side looks much more feasible to break apart.
Figure 5-14. A monolithic application with a high number of component dependencies
takes more effort to break apart (basketball sizing)
114
|
Chapter 5: Component-Based Decomposition Patterns
With this level of tight coupling between components, the answers to the three key
questions are not very encouraging:
1. Is it feasible to break apart the existing monolithic application? Maybe…
2. What is the rough overall level of effort for this migration? A basketball (much
harder)
3. Is this going to be a rewrite of the code or a refactoring of the code? Likely a com‐
bination of some refactoring and some rewriting of the existing code
Finally, consider the dependency diagram illustrated in Figure 5-15. In this case, the
architect should turn around and run in the opposite direction as fast as they can!
Figure 5-15. A monolithic application with too many component dependencies is not
feasible to break apart (airliner sizing)
Determine Component Dependencies Pattern
|
115
The answers to the three key questions for applications with this sort of component
dependency matrix are not surprising:
1. Is it feasible to break apart the existing monolithic application? No
2. What is the rough overall level of effort for this migration? An airliner
3. Is this going to be a rewrite of the code or a refactoring of the code? Total rewrite
of the application
We cannot stress enough the importance of these kinds of visual diagrams when
breaking apart a monolithic application. In essence these diagrams form a radar from
which to determine where the enemy (high component coupling) is located, and also
paint a picture of what the resulting service dependency matrix might look like if the
monolithic application were to be broken into a highly distributed architecture.
It has been our experience that component coupling is one of the most significant
factors in determining the overall success (and feasibility) of a monolithic migration
effort. Identifying and understanding the level of component coupling not only
allows the architect to determine the feasibility of the migration effort, but also what
to expect in terms of the overall level of effort. Unfortunately, all too often we see
teams jump straight into breaking a monolithic application into microservices
without having any analysis or visuals into what the monolithic application even
looks like. And not surprisingly, those teams struggle to break apart their monolithic
applications.
This pattern is useful not only for identifying the overall level of component coupling
in an application, but also for determining dependency refactoring opportunities
prior to breaking apart the application. When analyzing the coupling level between
components, it is important to analyze both afferent (incoming) coupling (denoted in
most tools as CA), and efferent (outgoing) coupling (denoted in most tools as CE).
CT, or total coupling, is the sum of both afferent and efferent coupling.
Many times, breaking apart a component can reduce the level of coupling of that
component. For example, assume component A has an afferent coupling level of 20
(meaning, 20 other components are dependent on the functionality of the compo‐
nent). This does not necessarily mean that all 20 of the other components require all
of the functionality from component A. Maybe 14 of the other components require
only a small part of the functionality contained in component A. Breaking compo‐
nent A into two different components (component A1 containing the smaller, cou‐
pled functionality, and component A2 containing the majority of the functionality)
reduces the afferent coupling in component A2 to 6, with component A1 having an
afferent coupling level of 14.
116
|
Chapter 5: Component-Based Decomposition Patterns
Fitness Functions for Governance
Two ways to automate the governance for component dependencies are to make sure
no component has “too many” dependencies, and to restrict certain components
from being coupled to other components. The fitness functions described next are
some ways of governing these type of dependencies.
Fitness function: No component shall have more than <some number> of total dependencies
This automated holistic fitness function can be triggered on deployment through a
CI/CD pipeline to make sure that the coupling level of any given component doesn’t
exceed a certain threshold. It is up to the architect to determine that this maximum
threshold should be based on the overall level of coupling within the application and
the number of components. An alert generated from this fitness function allows the
architect to discuss any sort of increase in coupling with the development team, pos‐
sibly promoting action to break apart components to reduce coupling. This fitness
function could also be modified to generate an alert for a threshold limit of incoming
only, outgoing only, or both (as separate fitness functions). Example 5-8 shows the
pseudocode for sending an alert if the total coupling (incoming and outgoing)
exceeds a combined level of 15, which for most applications would be considered rel‐
atively high.
Example 5-8. Pseudocode for limiting the total number of dependencies of any given
component
# Walk the directory structure, gathering components and the source code files
# contained within those components
LIST component_list = identify_components(root_directory)
MAP component_source_file_map
FOREACH component IN component_list {
LIST component_source_file_list = get_source_files(component)
ADD component, component_source_file_list TO component_source_file_map
}
# Determine how many references exist for each source file and send an alert if
# the total dependency count is greater than 15
FOREACH component,component_source_file_list IN component_source_file_map {
FOREACH source_file IN component_source_file_list {
incoming count = used_by_other_components(source_file, component_source_file_map) {
outgoing_count = uses_other_components(source_file) {
total_count = incoming count + outgoing count
}
IF total_count > 15 {
send_alert(component, total_count)
}
}
Determine Component Dependencies Pattern
|
117
Fitness function: <some component> should not have a dependency on <another
component>
This automated holistic fitness function can be triggered on deployment through a
CI/CD pipeline to restrict certain components from having a dependency on other
ones. In most cases, there will be one fitness function for each dependency restriction
so that, if there were 10 different component restrictions, there would be 10 different
fitness functions, one for each component in question. Example 5-9 shows an exam‐
ple using ArchUnit for ensuring that the Ticket Maintenance component
(ss.ticket.maintenance) does not have a dependency on the Expert Profile compo‐
nent (ss.expert.profile).
Example 5-9. ArchUnit code for governing dependency restrictions between components
public void ticket_maintenance_cannot_access_expert_profile() {
noClasses().that()
.resideInAPackage("..ss.ticket.maintenance..")
.should().accessClassesThat()
.resideInAPackage("..ss.expert.profile..")
.check(myClasses);
}
Sysops Squad Saga: Identifying Component Dependencies
Monday, November 15, 09:45
After reading about the Determine Component Dependencies pattern, Addison wondered what the Sysops Squad application dependency matrix looked like and whether it
was feasible to even break the application apart. Addison used an IDE plug-in to generate a component dependency diagram of the current Sysops Squad application. Initially,
Addison felt a bit discouraged because Figure 5-16 showed a lot of dependencies
between the Sysops Squad application components.
118
|
Chapter 5: Component-Based Decomposition Patterns
Figure 5-16. Component dependencies in the Sysops Squad application
However, after further analysis, Addison saw that the Notification component had the most dependencies, which was not surprising given that it’s a shared component. However, Addison also saw
lots of dependencies within the Ticketing and Reporting components. Both of these domain areas
have a specific component for shared code (interfaces, helper classes, entity classes, and so on). Realizing that both the ticketing and reporting shared code contains mostly compile-based class references and would likely be implemented as shared libraries rather than services, Addison filtered out
these components to get a better view of the dependencies between the core functionality of the
application, which is illustrated in Figure 5-17.
Determine Component Dependencies Pattern
|
119
Figure 5-17. Component dependencies in the Sysops Squad application without shared
library dependencies
With the shared components filtered out, Addison saw that the dependencies were fairly minimal.
Addison showed these results to Austen, and they both agreed that most of the components were
relatively self-contained and it appeared that the Sysops Squad application was a good candidate
for breaking apart into a distributed architecture.
Create Component Domains Pattern
While each component identified within a monolithic application can be considered
a possible candidate for a separate service, in most cases the relationship between a
service and components is a one-to-many relationship—that is, a single service may
contain one or more components. The purpose of the Create Component Domains
pattern is to logically group components together so that more coarse-grained
domain services can be created when breaking up an application.
120
|
Chapter 5: Component-Based Decomposition Patterns
Pattern Description
Identifying component domains—the grouping of components that perform some
sort of related functionality—is a critical part of breaking apart any monolithic appli‐
cation. Recall the advice from Chapter 4:
When breaking apart monolithic applications, consider first moving to service-based
architecture as a stepping-stone to other distributed architectures.
Creating component domains is an effective way of determining what will eventually
become domain services in a service-based architecture.
Component domains are physically manifested in an application through component
namespaces (or directories). Because namespace nodes are hierarchical in nature,
they become an excellent way of representing the domains and subdomains of func‐
tionality. This technique is illustrated in Figure 5-18, where the second node in the
namespace (.customer) refers to the domain, the third node represents a subdomain
under the customer domain (.billing), and the leaf node (.payment) refers to the
component. The .MonthlyBilling at the end of this namespace refers to a class file
contained within the Payment component.
Figure 5-18. Component domains are identified through the namespace nodes
Since many older monolithic applications were implemented prior to the widespread
use of domain-driven design, in many cases refactoring of the namespaces is needed
to structurally identify domains within the application. For example, consider the
components listed in Table 5-13 that make up the Customer domain within the
Sysops Squad application.
Table 5-13. Components related to the Customer domain before refactoring
Component
Namespace
Billing Payment
ss.billing.payment
Billing History
ss.billing.history
Customer Profile
ss.customer.profile
Support Contract
ss.supportcontract
Create Component Domains Pattern
|
121
Notice how each component is related to customer functionality, but the correspond‐
ing namespaces don’t reflect that association. To properly identify the Customer
domain (manifested through the namespace ss.customer), the namespaces for the
Billing Payment, Billing History, and Support Contract components would have to be
modified to add the .customer node at the beginning of the namespace, as shown in
Table 5-14.
Table 5-14. Components related to the Customer domain after refactoring
Component
Namespace
Billing Payment
ss.customer.billing.payment
Billing History
ss.customer.billing.history
Customer Profile
ss.customer.profile
Support Contract
ss.customer.supportcontract
Notice in the prior table that all of the customer-related functionality (billing, profile
maintenance, and support contract maintenance) is now grouped under .customer,
aligning each component with that particular domain.
Fitness Functions for Governance
Once refactored, it’s important to govern the component domains to ensure that
namespace rules are enforced and that no code exists outside the context of a compo‐
nent domain or subdomain. The following automated fitness function can be used to
help govern component domains once they are established within the monolithic
application.
Fitness function: All namespaces under <root namespace node> should be restricted to <list
of domains>
This automated holistic fitness function can be triggered on deployment through a
CI/CD pipeline to restrict the domains contained within an application. This fitness
function helps prevent additional domains from being inadvertently created by devel‐
opment teams and alerts the architect if any new namespaces (or directories) are cre‐
ated outside the approved list of domains. Example 5-10 shows an example using
ArchUnit for ensuring that only the ticket, customer, and admin domains exist within
an application.
122
|
Chapter 5: Component-Based Decomposition Patterns
Example 5-10. ArchUnit code for governing domains within an application
public void restrict_domains() {
classes()
.should().resideInAPackage("..ss.ticket..")
.orShould().resideInAPackage("..ss.customer..")
.orShould().resideInAPackage("..ss.admin..")
.check(myClasses);
}
Sysops Squad Saga: Creating Component Domains
Thursday, November 18, 13:15
Addison and Austen consulted with Parker, the Sysops Squad product owner, and
together they identified five main domains within the application: a Ticketing domain
(ss.ticket) containing all ticket-related functionality, including ticket processing, customer surveys, and knowledge base (KB) functionality; a Reporting domain (ss.report
ing) containing all reporting functionality; a Customer domain (ss.customer)
containing customer profile, billing, and support contracts; an Admin domain
(ss.admin) containing maintenance of users and Sysops Squad experts; and finally, a
Shared domain (ss.shared) containing login and notification functionality used by the other
domains.
Addison created a domain diagram (see Figure 5-19) showing the various domains and corresponding groups of components within each domain, and was satisfied with this grouping as no component was left out, and there was good cohesion between the components within each domain.
The exercise Addison did in diagramming and grouping the components was an important one as it
validated the identified domain candidates and also demonstrated the need for collaboration with
business stakeholders (such as the product owner or business application sponsor). Had the components not lined up properly or Addison was left with components that didn’t belong anywhere,
more collaboration with Parker (the product owner) would have been necessary.
Satisfied that all of the components fit nicely into these domains, Addison then looked at the various
component namespaces in Table 5-12 after applying the “Flatten Components Pattern” on page 101
and identified the component domain refactoring that needed to take place.
Create Component Domains Pattern
|
123
Figure 5-19. The five domains identified (with darkened borders) within the Sysops
Squad application
Addison started with the Ticket domain and saw that while the core ticket functionality started with
the namespace ss.ticket, the survey and knowledge base components did not. Therefore, Addison wrote an architecture story to refactor the components listed in Table 5-15 to align with the
ticketing domain.
124
| Chapter 5: Component-Based Decomposition Patterns
Table 5-15. Sysops Squad component refactoring for the Ticket domain
Component
Domain
Current namespace
Target namespace
KB Maint
Ticket
ss.kb.maintenance
ss.ticket.kb.maintenance
KB Search
Ticket
ss.kb.search
ss.ticket.kb.search
Ticket Shared
Ticket
ss.ticket.shared
Same (no change)
Ticket Maintenance
Ticket
ss.ticket.maintenance
Same (no change)
Ticket Completion
Ticket
ss.ticket.completion
Same (no change)
Ticket Assign
Ticket
ss.ticket.assign
Same (no change)
Ticket Route
Ticket
ss.ticket.route
Same (no change)
Survey
Ticket
ss.survey
ss.ticket.survey
Next Addison considered the customer-related components, and found that the billing and survey
components needed to be refactored to include them under the Customer domain, creating a Billing subdomain in the process. Addison wrote an architecture story for the refactoring of the Customer domain functionality, shown in Table 5-16.
Table 5-16. Sysops Squad component refactoring for the Customer domain
Component
Domain
Current namespace
Target namespace
Billing Payment
Customer
ss.billing.payment
ss.customer.billing.payment
Billing History
Customer
ss.billing.history
ss.customer.billing.history
Customer Profile
Customer
ss.customer.profile
Same (no change)
Support Contract
Customer
ss.supportcontract
ss.customer.supportcontract
By applying the “Identify and Size Components Pattern” on page 84, Addison found that the reporting domain was already aligned, and no further action was needed with the reporting components
listed in Table 5-17.
Table 5-17. Sysops Squad Reporting components are already aligned with the Reporting
domain
Component
Domain
Current namespace
Target namespace
Reporting Shared
Reporting
ss.reporting.shared
Same (no change)
Ticket Reports
Reporting
ss.reporting.tickets
Same (no change)
Expert Reports
Reporting
ss.reporting.experts
Same (no change)
Financial Reports
Reporting
ss.reporting.financial
Same (no change)
Create Component Domains Pattern
|
125
Addison saw that both the Admin and Shared domains needed alignment as well, and decided to
create a single architecture story for this refactoring effort and listed these components in
Table 5-18. Addison also decided to rename the ss.expert.profile namespace to ss.experts to
avoid an unnecessary Expert subdomain under the Admin domain.
Table 5-18. Sysops Squad component refactoring for the Admin and Shared domains
Component
Domain
Current namespace
Target namespace
Login
Shared
ss.login
aa.shared.login
Notification
Shared
ss.notification
ss.shared.notification
Expert Profile
Admin
ss.expert.profile
ss.admin.experts
User Maintenance
Admin
ss.users
ss.admin.users
With this pattern complete, Addison realized they were now prepared to structurally break apart the
monolithic application and move to the first stage of a distributed architecture by applying the Create Domain Services pattern (described in the next section).
Create Domain Services Pattern
Once components have been properly sized, flattened, and grouped into domains,
those domains can then be moved to separately deployed domain services, creating
what is known as a service-based architecture (see Appendix A). Domain services are
coarse-grained, separately deployed units of software containing all of the functional‐
ity for a particular domain (such as Ticketing, Customer, Reporting, and so on).
Pattern Description
The previous “Create Component Domains Pattern” on page 120 forms well-defined
component domains within a monolithic application and manifests those domains
through the component namespaces (or directory structures). This pattern takes
those well-defined component domains and extracts those component groups into
separately deployed services, known as a domain services, thus creating a servicebased architecture.
In its simplest form, service-based architecture consists of a user interface that
remotely accesses coarse-grained domain services, all sharing a single monolithic
database. Although there are many topologies within service-based architecture (such
as breaking up the user interface, breaking up the database, adding an API gateway,
and so on), the basic topology shown in Figure 5-20 is a good starting point for
migrating a monolithic application.
126
|
Chapter 5: Component-Based Decomposition Patterns
Figure 5-20. The basic topology for a service-based architecture
In addition to the benefits mentioned in “Component-Based Decomposition” on page
71, moving to service-based architecture first allows the architect and development
team to learn more about each domain service to determine whether it should be
broken into smaller services within a microservices architecture or left as a larger
domain service. Too many teams make the mistake of starting out too fine-grained,
and as a result must embrace all of the trappings of microservices (such as data
decomposition, distributed workflows, distributed transactions, operational automa‐
tion, containerization, and so on) without the need for all of those fine-grained
microservices.
Figure 5-21 illustrates how the Create Domain Services pattern works. Notice in the
diagram how the Reporting component domain defined in the “Create Component
Domains Pattern” on page 120 is extracted from of the monolithic application, form‐
ing its own separately deployed Reporting service.
Create Domain Services Pattern
|
127
Figure 5-21. Component domains are moved to external domain services
A word of advice, however: don’t apply this pattern until all of the component
domains have been identified and refactored. This helps reduce the amount of modi‐
fication needed to each domain service when moving components (and hence source
code) around. For example, suppose all of the ticketing and knowledge base function‐
ality in the Sysops Squad application was grouped and refactored into a Ticket
domain, and a new Ticket service created from that domain. Now suppose that the
customer survey component (identified through the ss.customer.survey name‐
space) was deemed part of the Ticket domain. Since the Ticket domain had already
been migrated, the Ticket service would now have to be modified to include the Sur‐
vey component. Better to align and refactor all of the components into component
domains first, then start migrating those component domains to domain services.
128
| Chapter 5: Component-Based Decomposition Patterns
Fitness Functions for Governance
It is important to keep the components within each domain service aligned with the
domain, particularly if the domain service will be broken into smaller microservices.
This type of governance helps keep domain services from becoming their own
unstructured monolithic service. The following fitness function ensures that the
namespace (and hence components) are kept consistent within a domain service.
Fitness function: All components in <some domain service> should start with the same
namespace
This automated holistic fitness function can be triggered on deployment through a
CI/CD pipeline to make sure the namespaces for components within a domain ser‐
vice remain consistent. For example, all components within the Ticket domain ser‐
vice should start with ss.ticket. Example 5-11 uses ArchUnit for ensuring this
constraint. Each domain service would have its own corresponding fitness function
based on its particular domain.
Example 5-11. ArchUnit code for governing components within the Ticket domain
service
public void restrict_domain_within_ticket_service() {
classes().should().resideInAPackage("..ss.ticket..")
.check(myClasses);
}
Sysops Squad Saga: Creating Domain Services
Tuesday, November 23, 09:04
Addison and Austen worked closely with the Sysops Squad development team to
develop a migration plan to stage the migration from component domains to domain
services. They realized this effort not only required the code within each component
domain to be extracted from the monolith and moved to a new project workspace, but
also for the user interface to now remotely access the functionality within that domain.
Working from the component domains identified previously in Figure 5-19, the team
migrated each component, one at a time, eventually arriving at a service-based architecture, as shown in Figure 5-22. Notice how each domain area identified in the previous pattern now
becomes a separately deployed service.
Create Domain Services Pattern
|
129
Figure 5-22. Separately deployed domain services result in a distributed Sysops Squad
application
Summary
It has been our experience that “seat-of-the-pants” migration efforts rarely produce
positive results. Applying these component-based decomposition patterns provides a
structured, controlled, and incremental approach for breaking apart monolithic
architectures. Once these patterns are applied, teams can now work to decompose
monolithic data (see Chapter 6) and begin breaking apart domain services into more
fine-grained microservices (see Chapter 7) as needed.
130
|
Chapter 5: Component-Based Decomposition Patterns
CHAPTER 6
Pulling Apart Operational Data
Thursday, October 7, 08:55
Now that the Sysops Squad application was successfully broken into separately
deployed domain services, Addison and Austen both realized that it was time to start
thinking about breaking apart the monolithic Sysops Squad database. Addison agreed
to start this effort, while Austen began to work on enhancing the CI/CD deployment
pipeline. Addison met with Dana, the Sysops Squad data architect, and also Devon, one
of the DBAs supporting the Penultimate Electronics databases.
“I’d like your opinions on how we might go about breaking up the Sysops Squad database,” said Addison.
“Wait a minute,” said Dana. “Who said anything about breaking apart the database?”
“Addison and I agreed last week that we needed to break up the Sysops Squad database,” said
Devon. “As you know, the Sysops Squad application has been going through a major overhaul, and
breaking apart the data is part of that overhaul.”
“I think the monolithic database is just fine,” said Dana. “I see no reason why it should be broken
apart. Unless you can convince me otherwise, I’m not going to budge on this issue. Besides, do you
know how hard it would be to break apart that database?”
“Of course it will be difficult,” said Devon, “but I know of a five-step process leveraging what are
known as data domains that would work really well on this database. That way, we can even start
investigating using different kinds of databases for certain parts of the application, like the knowledge base and even the customer survey functionality.”
“Let’s not get ahead of ourselves,” said Dana. “And let’s also not forget that I am the one who is
responsible for all of these databases.”
131
Addison quickly realized things were spiraling out of control, and quickly put some key negotiation
and facilitation skills to use. “OK,” said Addison, “we should have included you in our initial discussions, and for that I apologize. I should have known better. What can we do to bring you on board
and help us decompose the Sysops Squad database?”
“That’s easy,” said Dana. “Convince me that the Sysops Squad database really does need to be broken
apart. Provide me with a solid justification. If you can do that, then we’ll talk about Devon’s five-step
process. Otherwise, it stays as it is.”
Breaking apart a database is hard—much harder, in fact, than breaking apart applica‐
tion functionality. Because data is generally the most important asset in the company,
there is greater risk of business and application disruption when breaking apart or
restructuring data. Also, data tends to be highly coupled to application functionality,
making it harder to identify well-defined seams within a large data model.
In the same way a monolithic application is broken into separate deployment units,
there are times when it is desirable (or even necessary) to break up a monolithic data‐
base as well. Some architecture styles, such as microservices, require data to be broken
apart to form well-defined bounded contexts (where each service owns its own data),
whereas other distributed architectures, such as service-based architecture, allow
services to share a single database.
Interestingly enough, some of the same techniques used to break apart application
functionality can be applied to breaking apart data as well. For example, components
translate to data domains, class files translate to database tables, and coupling points
between classes translate to database artifacts such as foreign keys, views, triggers, or
even stored procedures.
In this chapter, we explore some of the drivers for decomposing data and show tech‐
niques for how to effectively break apart monolithic data into separate data domains,
schemas, and even separate databases in an iterative and controlled fashion. Knowing
that the database world is not all relational, we also discuss various types of databases
(relational, graph, document, key-value, columnar, NewSQL, and cloud native) and
outline the various trade-offs associated with each of these database types.
Data Decomposition Drivers
Breaking apart a monolithic database can be a daunting task, and as such it’s impor‐
tant to understand if (and when) a database should be decomposed, as illustrated in
Figure 6-1. Architects can justify a data decomposition effort by understanding and
analyzing data disintegrators (drivers that justify breaking apart data) and data inte‐
grators (drivers that justify keeping data together). Striving for a balance between
132
| Chapter 6: Pulling Apart Operational Data
these two driving forces and analyzing the trade-offs of each is the key to getting data
granularity right.
Figure 6-1. Under what circumstances should a monolithic database be decomposed?
In this section, we will explore the data disintegrators and data integrators used to
help make the right choice when considering breaking apart monolithic data.
Data Disintegrators
Data disintegration drivers provide answers and justifications for the question “when
should I consider breaking apart my data?” The six main disintegration drivers for
breaking apart data include the following:
Change control
How many services are impacted by a database table change?
Connection management
Can my database handle the connections needed from multiple distributed serv‐
ices?
Scalability
Can the database scale to meet the demands of the services accessing it?
Fault tolerance
How many services are impacted by a database crash or maintenance downtime?
Architectural quanta
Is a single shared database forcing me into an undesirable single architecture
quantum?
Database type optimization
Can I optimize my data by using multiple database types?
Data Decomposition Drivers
|
133
Each of these disintegration drivers is discussed in detail in the following sections.
Change control
One of the primary data disintegration drivers is controlling changes in the database
table schemas. Dropping tables or columns, changing table or column names, and
even changing the column type in a table break the corresponding SQL accessing
those tables, and consequently break corresponding services using those tables. We
call these types of changes breaking changes as opposed to adding tables or columns
in a database, which generally do not impact existing queries or writes. Not surpris‐
ingly, change control is most impacted when using relational databases, but other
database types can create change control issues as well (see “Selecting a Database
Type” on page 161).
As illustrated in Figure 6-2, when breaking changes occur to a database, multiple
services must be updated, tested, and deployed together with the database changes.
This coordination can quickly become both difficult and error prone as the number
of separately deployed services sharing the same database increases. Imagine trying to
coordinate 42 separately deployed services for a single breaking database change!
Figure 6-2. Services impacted by the database change must be deployed together with the
database
Coordinating changes to multiple distributed services for a shared database change is
only half the story. The real danger of changing a shared database in any distributed
architecture is forgetting about services that access the table just changed. As illustra‐
ted in Figure 6-3, those services become nonoperational in production until they can
be changed, tested, and redeployed.
134
|
Chapter 6: Pulling Apart Operational Data
Figure 6-3. Services impacted by a database change but forgotten will continue to fail
until redeployed
In most applications, the danger of forgotten services is mitigated by diligent impact
analysis and agressive regression testing. However, consider a microservices ecosys‐
tem with 400 services, all sharing the same monolithic highly available clustered rela‐
tional database. Imagine running around to all the development teams in many
domain areas, trying to find out which services use the table being changed. Also
imagine having to then coordinate, test, and deploy all of these services together as a
single unit, along with the database. Thinking about this scenario starts to become a
mind-numbing exercise, usually leading to some degree of insanity.
Breaking apart a database into well-defined bounded contexts significantly helps con‐
trol breaking database changes. The bounded context concept comes from the semi‐
nal book Domain-Driven Design by Eric Evans (Addison-Wesley) and describes the
source code, business logic, data structures, and data all bound together—encapsula‐
ted—within a specific context. As illustrated in Figure 6-4, well-formed bounded con‐
texts around services and their corresponding data helps control change, because
change is isolated to just those services within that bounded context.
Most typically, bounded contexts are formed around services and the data the serv‐
ices owns. By “own” we mean a service that writes to the database (as opposed to hav‐
ing read-only access to the data). We discuss distributed data ownership in more
detail in Chapter 9.
Data Decomposition Drivers
|
135
Figure 6-4. Database changes are isolated to only those services within the associated
bounded context
Notice in Figure 6-4 that Service C needs access to some of the data in Database D
that is contained in a bounded context with Service D. Since Database D is in a differ‐
ent bounded context, Service C cannot directly access the data. This would not only
violate the bounded context rule, but also create a mess with regard to change con‐
trol. Therefore, Service C must ask Service D for the data. There are many ways of
accessing data a service doesn’t own while still maintaining a bounded context. These
techniques are discussed in detail in Chapter 10.
One important aspect of a bounded context related to the scenario between Service C
needing data and Service D owning that data within its bounded context is that of
database abstraction. Notice in Figure 6-5 that Service D is sending data that was
requested by Service C through some sort of contract (such as JSON, XML, or maybe
even an object).
The advantage of the bounded context is that the data sent to Service C can be a dif‐
ferent contract than the schema for Database D. This means that a breaking change to
some table in Database D impacts only Service D and not necessarily the contract of
the data sent to Service C. In other words, Service C is abstracted from the actual
schema structure of Database D.
136
|
Chapter 6: Pulling Apart Operational Data
Figure 6-5. The contract from a service call abstracts the caller from the underlying data‐
base schema
To illustrate the power of this bounded context abstraction within a distributed archi‐
tecture, assume Database D has a Wishlist table with the following structure:
CREATE TABLE Wishlist
(
CUSTOMER_ID VARCHAR(10),
ITEM_ID VARCHAR(20),
QUANTITY INT,
EXPIRATION_DT DATE
);
The corresponding JSON contract that Service D sends to Service C requesting wish
list items is as follows:
{
"$schema": "http://json-schema.org/draft-04/schema#",
"properties": {
"cust_id": {"type": "string"},
"item_id": {"type": "string"},
"qty": {"type": "number"},
"exp_dt": {"type": "number"}
},
}
Notice how the expiration data field (exp_dt) in the JSON schema is named differ‐
ently than the database column name and is specified as a number (a long value rep‐
resenting the epoch time—the number of milliseconds since midnight on 1 January
1970), whereas in the database it is represented as a DATE field. Any column name
Data Decomposition Drivers
|
137
change or column type change made in the database no longer impacts Service C
because of the separate JSON contract.
To illustrate this point, suppose the business decides to no longer expire wish list
items. This would require a change in the table structure of the database:
ALTER TABLE Wishlist
DROP COLUMN EXPIRATION_DT;
Service D would have to be modified to accommodate this change because it is within
the same bounded context as the database, but the corresponding contract would not
have to change at the same time. Until the contract is eventually changed, Service D
could either specify a date far into the future or set the value to zero indicating the
item doesn’t expire. The bottom line is that Service C is abstracted from breaking
changes made to Database D due to the bounded context.
Connection management
Establishing a connection to a database is an expensive operation. A database connec‐
tion pool is often used not only to increase performance, but also to limit the number
of concurrent connections an application is allowed to use. In monolithic applica‐
tions, the database connection pool is usually owned by the application (or applica‐
tion server). However, in distributed architectures, each service—or more specifically,
each service instance—typically has its own connection pool. As illustrated in Figure
6-6, when multiple services share the same database, the number of connections can
quickly become saturated, particularly as the number of services or service instances
increase.
Figure 6-6. Database connections can quickly get saturated with multiple service
instances
138
|
Chapter 6: Pulling Apart Operational Data
Reaching (or exceeding) the maximum number of available database connections is
yet another driver to consider when deciding whether to break apart a database. Fre‐
quent connection waits (the amount of time it takes waiting for a connection to
become available) is usually the first sign that the maximum number of database con‐
nections has been reached. Since connection waits can also manifest themselves as
request time-outs or tripped circuit breakers, looking for connection waits is usually
the first thing we recommend if these conditions frequently occur when using a
shared database.
To illustrate the issues associated with database connections and distributed architec‐
ture, consider the following example: a monolithic application with 200 database con‐
nections is broken into a distributed architecture consisting of 50 services, each with
10 database connections in its connection pool.
Original monolithic application
Distributed services
Connections per service
Minimum service instances
Total service connections
200 connections
50
10
2
1,000
Notice how the number of database connections within the same application context
grew from 200 to 1,000, and the services haven’t even started scaling yet! Assuming
half of the services scale to an average of 5 instances each, the number of database
connections quickly grows to 1,700.
Without some sort of connection strategy or governance plan, services will try to use
as many connections as possible, frequently starving other services from much
needed connections. For this reason, it’s important to govern how database connec‐
tions are used in a distributed architecture. One effective approach is to assign each
service a connection quota to govern the distribution of available database connec‐
tions across services. A connection quota specifies the maximum number of database
connections a service is allowed to use or make available in its connection pool.
By specifying a connection quota, services are not allowed to create more database
connections than are allocated to it. If a service reaches the maximum number of
database connections in its quota, it must wait for one of the connections it’s using to
become available. This method can be implemented using two approaches: evenly
distributing the same connection quota to every service, or assigning a different con‐
nection quota to each service based on its needs.
The even distribution approach is typically used when first deploying services, and it
is not known yet how many connections each service will need during normal and
peak operations. While simple, this approach is not overly efficient because some
Data Decomposition Drivers
|
139
services may need more connections than others, while some connections held by
other services may go unused.
While more complex, the variable distribution approach is much more efficient for
managing database connections to a shared database. With this approach, each ser‐
vice is assigned a different connection quota based on its functionality and scalability
requirements. The advantage of this approach is that it optimizes the use of available
database connections across distributed services, making sure those services that
require more database connections have them available for use. However, the disad‐
vantage is that it requires knowledge about the nature of the functionality and the
scalability requirements of each service.
We usually recommend starting out with the even distribution approach and creating
fitness functions to measure the concurrent connection usage for each service. We
also recommend keeping the connection quota values in an external configuration
server (or service) so that the values can be easily adjusted either manually or pro‐
grammatically through simple machine learning algorithms. This technique not only
helps mitigate connection saturation risk, but also properly balances available data‐
base connections between distributed services to ensure that no idle connections are
wasted.
Table 6-1 shows an example of starting out using the even distribution approach for a
database that can support a maximum of 100 concurrent connections. Notice that
Service A has only ever needed a maximum of 5 connections, Service C only 15 con‐
nections, and Service E only 14 connections, whereas Service B and Service D have
reached their max connection quota and have experienced connection waits.
Table 6-1. Connection quota allocations evenly distributed
→
→
Service
Quota
Max used
Waits
A
B
C
D
E
20
20
20
20
20
5
20
15
20
14
No
Yes
No
Yes
No
Since Service A is well below its connection quota, this is a good place to start reallo‐
cating connections to other services. Moving five database connections to Service B
and five database connections to Service D yields the results shown in Table 6-2.
140
| Chapter 6: Pulling Apart Operational Data
Table 6-2. Connection quota allocations with varying distributions
→
Service
Quota
Max used
Waits
A
B
C
D
E
10
25
20
25
20
5
25
15
25
14
No
Yes
No
No
No
This is better, but Service B is still experiencing connection waits, indicating that it
requires more connections than it has in its connection quota. Readjusting the quotas
even further by taking two connections each from Service A and Service E yields
much better results, as shown in Table 6-3.
Table 6-3. Further connection quota tuning results in no connection waits
Service
Quota
Max used
Waits
A
B
C
D
E
8
29
20
25
18
5
27
15
25
14
No
No
No
No
No
This analysis, which can be derived from continuous fitness functions that gather
streamed metrics data from each service, can also be used to determine how close the
maximum number of connections used is to the maximum number of connections
available, and also how much buffer exists for each service in terms of its quota and
maximum connections used.
Scalability
One of the many advantages of a distributed architecture is scalability—the ability for
services to handle increases in request volume while maintaining a consistent
response time. Most cloud-based and on-prem infrastructure-related products do a
good job at ensuring that services, containers, HTTP servers, and virtual machines
scale to satisfy increases in demand. But what about the database?
As illustrated in Figure 6-7, service scalability can put a tremendous strain on the
database, not only in terms of database connections (as discussed in the prior sec‐
tion), but also on throughput and database capacity. In order for a distributed system
to scale, all parts of the system need to scale—including the database.
Data Decomposition Drivers
|
141
Figure 6-7. The database must also scale when services scale
Scalability is another data disintegration driver to consider when thinking about
breaking apart a database. Database connections, capacity, throughput, and perfor‐
mance are all factors in determining whether a shared database can meet the
demands of multiple services within a distributed architecture.
Consider the refined variable database connection quotas in Table 6-3 in the prior
section. When services scale by adding multiple instances, the picture changes dra‐
matically, as shown in Table 6-4, where the total number of database connections is
100.
Table 6-4. When services scale, more connection are used than are available
Service
Quota
Max used
Instances
Total used
A
B
C
D
E
TOTAL
8
29
20
25
18
100
5
27
15
25
14
86
2
3
3
2
4
14
10
81
45
50
56
242
142
|
Chapter 6: Pulling Apart Operational Data
Notice that even though the connection quota is distributed to match the 100 data‐
base connections available, once services start to scale, the quota is no longer valid
because the total number of connections used increases to 242, which is 142 more
connections than are available in the database. This will likely result in connection
waits, which in turn will result in overall performance degradation and request
time-outs.
Breaking data into separate data domains or even a database-per-service, as illustra‐
ted in Figure 6-8, requires fewer connections to each database, hence providing better
database scalability and performance as the services scale.
Figure 6-8. Breaking apart the database provides better database scalability
In addition to database connections, another factor to consider with respect to scala‐
bility is the load placed on the database. By breaking apart a database, less load is
placed on each database, thereby also improving overall performance and scalability.
Fault tolerance
When multiple services share the same database, the overall system becomes less fault
tolerant because the database becomes a single point of failure (SPOF). Here, we are
defining fault tolerance as the ability of some parts of the system to continue uninter‐
rupted when a service or database fails. Notice in Figure 6-9 that when sharing a sin‐
gle database, overall fault tolerance is low because if the database goes down, all
services become nonoperational.
Data Decomposition Drivers
|
143
Figure 6-9. If the database goes down, all services become nonoperational
Fault tolerance is another driver for considering breaking apart data. If fault tolerance
is required for certain parts of the system, breaking apart the data can remove the sin‐
gle point of failure in the system, as shown in Figure 6-10. This ensures that some
parts of the system are still operational in the event of a database crash.
Figure 6-10. Breaking apart the database achieves better fault tolerance
Notice that since the data is now broken apart, if Database B goes down, only Service
B and Service C are impacted and become nonoperational, whereas the other services
continue to operate uninterrupted.
Architectural quantum
Recall from Chapter 2 that an architectural quantum is defined as an independently
deployable artifact with high functional cohesion, high static coupling, and synchro‐
nous dynamic coupling. The architecture quantum helps provide guidance in terms
of when to break apart a database, making it another data disintegration driver.
144
| Chapter 6: Pulling Apart Operational Data
Consider the services in Figure 6-11, where Service A and Service B require different
architectural characteristics than the other services. Notice in the diagram that
although Service A and Service B are grouped together, they do not form a separate
quantum from the other services because of a single shared database. Thus, all five
services, along with the database, form a single architectural quantum.
Figure 6-11. The database is part of the architectural quantum
Because the database is included in the functional cohesion part of the architecture
quantum definition, it is necessary to break apart the data so that each resulting part
can be in its own quantum. Notice in Figure 6-12 that since the database is broken
apart, Service A and Service B, along with the corresponding data, are now a separate
quantum from the one formed with services C, D, and E.
Data Decomposition Drivers
|
145
Figure 6-12. Breaking up the database forms two architectural quanta
Database type optimization
It’s often the case that not all data is treated the same. When using a monolithic data‐
base, all data must adhere to that database type, therefore producing potentially suboptimal solutions for certain types of data.
Breaking apart monolithic data allows the architect to move certain data to a more
optimal database type. For example, suppose a monolithic relational database stored
application-related transactional data, including reference data in the form of keyvalue pairs (such as country codes, product codes, warehouse codes, and so on). This
type of data is difficult to manage in a relational database because the data is not rela‐
tional in nature, but rather key-value. Hence, a key-value database (see “Key-Value
Databases” on page 165) would produce a more optimal solution than a relational
database.
Data Integrators
Data integrators do the exact opposite of the data disintegrators discussed in the prior
section. These drivers provide answers and justifications for the question “when
should I consider putting data back together?” Along with data disintegrators, data
integrators provide the balance and trade-offs for analyzing when to break apart data
and when not to.
146
| Chapter 6: Pulling Apart Operational Data
The two main integration drivers for pulling data back together are the following:
Data relationships
Are there foreign keys, triggers, or views that form close relationships between
the tables?
Database transactions
Is a single transactional unit of work necessary to ensure data integrity and con‐
sistency?
Each of these integration drivers is discussed in detail in the following sections.
Data relationships
Like components within an architecture, database tables can be coupled as well, par‐
ticularly with regard to relational databases. Artifacts like foreign keys, triggers,
views, and stored procedures tie tables together, making it difficult to pull data apart;
see Figure 6-13.
Imagine walking up to your DBA or data architect and telling them that since the
database must be broken apart to support tightly formed bounded contexts within a
microservices ecosystem, every foreign key and view in the database needs to be
removed! That’s not a likely (or even feasible) scenario, yet that is precisely what
would need to happen to support a database-per-service pattern in microservices.
Figure 6-13. Foreign keys (FK), triggers, and views create tightly coupled relationships
between data
Data Decomposition Drivers
|
147
These artifacts are necessary in most relational databases to support data consistency
and data integrity. In addition to these physical artifacts, data may also be logically
related, such as a problem ticket table and its corresponding problem ticket status
table. However, as illustrated in Figure 6-14, these artifacts must be removed when
moving data to another schema or database to form bounded contexts.
Figure 6-14. Data artifacts must be removed when breaking apart data
Notice that the foreign key (FK) relationship between the tables in Service A can be
preserved because the data is in the same bounded context, schema, or database.
However, the foreign keys (FK) between the tables in Service B and Service C must be
removed (as well as the view that is used in Service C) because those tables are associ‐
ated with different databases or schemas.
The relationship between data, either logical or physical, is a data integration driver,
thus creating a trade-off between data disintegrators and data integrators. For exam‐
ple, is change control (a data disintegrator) more important than preserving the for‐
eign key relationships between the tables (a data integrator)? Is fault tolerance (a data
disintegrator) more important than preserving materialized views between tables (a
data integrator)? Identifying what is more important helps make the decision about
whether the data should be broken apart and what the resulting schema granularity
should be.
Database transactions
Another data integrator is that of database transactions, something we discuss in
detail in “Distributed Transactions” on page 263. As shown in Figure 6-15, when a
single service does multiple database write actions to separate tables in the same data‐
base or schema, those updates can be done within an Atomicity, Consistency, Isola‐
tion, Durability (ACID) transaction and either committed or rolled back as a single
unit of work.
148
|
Chapter 6: Pulling Apart Operational Data
Figure 6-15. A single transactional unit of work exists when the data is together
However, when data is broken apart into either separate schemas or databases, as
illustrated in Figure 6-16, a single transactional unit of work no longer exists because
of the remote calls between services. This means that an insert or update can be com‐
mitted in one table, but not in the other tables because of error conditions, resulting
in data consistency and integrity issues.
Figure 6-16. Single unit of work transactions don’t exist when data is broken apart
While we dive into the details of distributed transaction management and transac‐
tional sagas in Chapter 12, the point here is to emphasize that database transactions
are yet another data integration driver, and should be taken into account when con‐
sidering breaking apart a database.
Data Decomposition Drivers
|
149
Sysops Squad Saga: Justifying Database Decomposition
Monday, November 15, 15:55
Armed with their justifications, Addison and Devon met to convince Dana that it was
necessary to break apart the monolithic Sysops Squad database.
“Hi, Dana,” said Addison. “We think we have enough evidence to convince you that it’s
necessary to break apart the Sysops Squad database.”
“I’m all ears,” said Dana, arms crossed and ready to argue that the database should
remain as is.
“I’ll start,” said Addison. “Notice how these logs continuously show that whenever the operational
reports run, the ticketing functionality in the application freezes up?”
“Yeah,” said Dana, “I’ll admit that even I suspected that. It’s clearly something wrong with the way the
ticketing functionality is accessing the database, not reporting.”
“Actually,” said Addison, “it’s a combination of both ticketing and reporting. Look here.”
Addison showed Dana metrics and logs that demonstrated some of the queries were necessarily
wrapped in threads, and that the queries from the ticketing functionality were timing out because of
a wait state when the reporting queries were run. Addison also showed how the reporting part of
the system used parallel threads to query parts of the more complex reports concurrently, essentially taking up all of the database connections.
“OK, I can see how having a separate reporting database would help the situation from a database
connection perspective. But that still doesn’t convince me that the nonreporting data should be
broken apart,” said Dana.
“Speaking of database connections,” said Devon, “look at this connection pool estimate as we start
breaking apart the domain services.”
Devon showed Dana the number of estimated services in the final planned Sysops Squad distributed application, including the projected number of instances for each of the services as the
application scales. Dana explained to Devon that the connection pool was contained within each
separate service instance, not like in the current phase of the migration where the application server
owned the connection pool.
“So you see, Dana,” said Devon, “with these projected estimates, we will need an additional 2,000
connections to the database to provide the scalability we need to handle the ticket load, and we
simply do not have them with a single database.”
Dana took a moment to look over the numbers. “Do you agree with these numbers, Addison?”
“I do,” said Addison. “Devon and I came up with them ourselves after a lot of analysis based on the
amount of HTTP traffic as well as the projected growth rates supplied by Parker.”
150
|
Chapter 6: Pulling Apart Operational Data
“I must admit,” said Dana, “this is good stuff you’ve both prepared. I particularly like that you’ve
already thought about not having services connect to multiple databases or schemas. As you know,
in my book that’s a no-go.”
“Us, too. However, we have one more justification to talk to you about,” said Addison. “As you may or
may not know, we’ve been having lots of issues with regard to the system not being available for our
customers. While breaking apart the services provides us with some level of fault tolerance, if a monolithic database should go down for either maintenance or a server crash, all services would become
nonoperational.”
“What Addison is saying,” added Devon, “is that by breaking apart the database, we can provide better fault tolerance by creating domain silos for the data. In other words, if the survey database were
to go down, ticketing functionality would still be available.”
“We call that an architectural quantum,” said Addison. “In other words, since the database is part of
the static coupling of a system, breaking it apart would make the core ticketing functionality standalone and not synchronously dependent on other parts of the system.”
“Listen,” said Dana, “you’ve convinced me that there’s good reasons to break apart the Sysops Squad
database, but explain to me how you can even think about doing that. Do you realize how many
foreign keys and views there are in that database? There’s no way you’re going to be able remove all
of those things.”
“We don’t necessarily have to remove all of those artifacts. That’s where data domains and the fivestep process come into play,” said Devon. “Here, let me explain…”
Decomposing Monolithic Data
Decomposing a monolithic database is hard, and requires an architect to collaborate
closely with the database team to safely and effectively break apart the data. One par‐
ticularly effective technique for breaking apart data is to leverage what is known as
the five-step process. As illustrated in Figure 6-17, this evolutionary and iterative pro‐
cess leverages the concept of a data domain as a vehicle for methodically migrating
data into separate schemas, and consequently different physical databases.
Figure 6-17. Five-step process for decomposing a monolithic database
Decomposing Monolithic Data
|
151
A data domain is a collection of coupled database artifacts—tables, views, foreign
keys, and triggers—that are all related to a particular domain and frequently used
together within a limited functional scope. To illustrate the concept of a data domain,
consider the Sysops Squad tables introduced in Table 1-2 and the corresponding pro‐
posed data domain assignments shown in Table 6-5.
Table 6-5. Existing Sysops Squad database tables assigned to data domains
Table
Proposed data domains
customer
customer_notification
survey
question
survey_administered
survey_question
survey_response
billing
contract
payment_method
payment
sysops_user
profile
expert_profile
expertise
location
article
tag
keyword
article_tag
article_keyword
ticket
ticket_type
ticket_history
Customer
Customer
Survey
Survey
Survey
Survey
Survey
Payment
Payment
Payment
Payment
Profile
Profile
Profile
Profile
Profile
Knowledge Base
Knowledge Base
Knowledge Base
Knowledge Base
Knowledge Base
Ticketing
Ticketing
Ticketing
Table 6-5 lists six data domains within the Sysops Squad application: Customer, Sur‐
vey, Payment, Profile, Knowledge base, and Ticketing. The billing table belongs to
the Payment data domain, ticket and ticket_type tables belong to the Ticketing
data domain, and so on.
152
|
Chapter 6: Pulling Apart Operational Data
One way to conceptually think about data domains is to think about the database as a
soccer ball, where each white hexagon represents a separate data domain. As illustra‐
ted in Figure 6-18, each white hexagon of the soccer ball contains a collection of
domain-related tables along with all of the coupling artifacts (such as foreign keys,
views, stored procedures, and so on).
Figure 6-18. Database objects in a hexagon belong in a data domain
Visualizing the database this way allows the architect and database team to clearly see
data domain boundaries and also the cross-domain dependencies (such as foreign
keys, views, stored procedures, and so on) that need to be broken. Notice in Figure
6-18 that within each white hexagon, all data table dependencies and relationships
can be preserved, but not between each white hexagon. For example, in the diagram
notice that solid lines represent dependencies that are self-contained to the data
domain, while the dotted lines cross data domains and must be removed when the
data domains are extracted into separate schemas.
Decomposing Monolithic Data
|
153
When extracting a data domain, these cross-domain dependencies must be removed.
This means removing foreign-key constraints, views, triggers, functions, and stored
procedures between data domains. Database teams can leverage the refactoring pat‐
terns found in the book Refactoring Databases: Evolutionary Database Design, by Scott
Ambler and Pramod Sadalage (Addison-Wesley), to safely and iteratively remove
these data dependencies.
Figure 6-19. Tables belonging to data domains, extracted out, and connections that need
to be broken
To illustrate the process of defining a data domain and removing cross-domain refer‐
ences, consider the diagram in Figure 6-19, where a data domain representing Pay‐
ment is created. Since the customer table belongs to a different data domain than the
v_customer_contract, the customer table must be removed from the view in the
Payment domain. The original view v_customer_contract prior to defining the data
domain is defined in Example 6-1.
154
| Chapter 6: Pulling Apart Operational Data
Example 6-1. Database view to get open tickets for customer with cross-domain joins
CREATE VIEW [payment].[v_customer_contract]
AS
SELECT
customer.customer_id, customer.customer_name,
contract.contract_start_date, contract.contract_duration,
billing.billing_date, billing.billing_amount
FROM payment.contract AS contract
INNER JOIN customer.customer AS customer
ON ( contract.customer_id = customer.customer_id )
INNER JOIN payment.billing AS billing
ON ( contract.contract_id = billing.contract_id )
WHERE contract.auto_renewal = 0
Notice in the updated view shown in Example 6-2 that the join between customer
and payment tables is removed, as is the column for the customer name (cus
tomer.customer_name).
Example 6-2. Database view to get open tickets in ticket domain for a given customer
CREATE VIEW [payment].[v_customer_contract]
AS
SELECT
billing.customer_id, contract.contract_start_date,
contract.contract_duration, billing.billing_date,
billing.billing_amount
FROM payment.contract AS contract
INNER JOIN payment.billing AS billing
ON ( contract.contract_id = billing.contract_id )
WHERE contract.auto_renewal = 0
The bounded context rules for data domains apply just the same as individual tables
—a service cannot talk to multiple data domains. Therefore, by removing this table
from the view, the Payment service must now call the Customer service to get the
customer name that it originally had from the view.
Once architects and database teams understand the concept of a data domain, they
can apply the five-step process for decomposing a monolithic database. Those five
steps are outlined in the following sections.
Decomposing Monolithic Data
|
155
Step 1: Analyze Database and Create Data Domains
As illustrated in Figure 6-20, all services have access to all data in the database. This
practice, known as the shared database integration style described by Gregor Hohpe
and Bobby Woolf in their book Enterprise Integration Patterns: Designing, Building,
and Deploying Messaging Solutions (Addison-Wesley), creates a tight coupling
between data and the services accessing that data. As discussed in “Data Decomposi‐
tion Drivers” on page 132, this tight coupling in the database makes change manage‐
ment very difficult.
Figure 6-20. Multiple services use the same database, accessing all the tables necessary
for read or write purposes
The first step in breaking apart a database is to identify specific domain groupings
within the database. For example, as shown in Table 6-5, related tables are grouped
together to help identify possible data domains.
Step 2: Assign Tables to Data Domains
The next step is to group tables along a specific bounded context, assigning tables
that belong to a specific data domain into their own schema. A schema is a logical
construct in database servers. A schema contain objects such as tables, views, func‐
tions, and so on. In some database servers, like Oracle, the schema is same as the user,
while in other databases, like SQL Server, a schema is logical space for database
objects where users have access to these schemas.
As illustrated in Figure 6-21, we have created schemas for each data domain and
moved tables to the schemas to which they belong.
156
|
Chapter 6: Pulling Apart Operational Data
Figure 6-21. Services use the primary schema according to their data domain needs
When tables belonging to different data domains are tightly coupled and related to
one another, data domains must necessarily be combined, creating a broader boun‐
ded context where multiple services own a specific data domain. Combining data
domains is discussed in more detail in Chapter 9.
Data Domain Versus Database Schema
A data domain is an architectural concept, whereas a schema is a database construct
that holds the database objects belonging to a particular data domain. While the rela‐
tionship between a data domain and a schema is usually one to one, data domains can
be mapped to one or more schemas, particularly when combining data domains
because of tightly coupled data relationships. We will be referring to a data domain
and schema to mean the same thing, and will be using the terms interchangeably.
To illustrate the assignment of tables to schemas, consider the Sysops Squad example
where the billing table must be moved from its original schema to another data
domain schema called payment:
ALTER SCHEMA payment TRANSFER sysops.billing;
Alternatively, a database team can create synonyms for tables that do not belong in
their schema. Synonyms are database constructs, similar to symlink, that provide an
alternate name for another database object that can exist in the same or different
schema or server. While the idea of synonyms is to eliminate cross-schema queries,
read or write privileges are needed to access them.
Decomposing Monolithic Data
|
157
To illustrate this practice, consider the following cross-domain query:
SELECT
history.ticket_id, history.notes, agent.name
FROM ticket.ticket_history AS history
INNER JOIN profile.sysops_user AS agent
ON ( history.assigned_to_sysops_user_id = agent.sysops_user_id )
Next, create a synonym for the profile.sysops_user table in the ticketing schema:
CREATE SYNONYM ticketing.sysops_user
FOR profile.sysops_user;
GO
As a result, the query can leverage the synonym sysops_user rather than the crossdomain table:
SELECT
history.ticket_id, history.notes, agent.name
FROM ticket.ticket_history AS history
INNER JOIN ticket.sysops_user AS agent
ON ( history.assigned_to_sysops_user_id = agent.sysops_user_id )
Unfortunately, creating synonyms this way for tables that are accessed across schemas
provides the application developers with coupling points. To form proper data
domains, these coupling points need to be broken apart at some later time, therefore
moving the integration points from the database layer to the application layer.
While synonyms do not really get rid of cross-schema queries, they do allow for eas‐
ier dependency checking and code analysis, making it easier to split these later on.
Step 3: Separate Database Connections to Data Domains
In this step, the database connection logic within each service is refactored to ensure
services connect to a specific schema and have read and write access to the tables
belonging only to their data domain. This transition, illustrated in Figure 6-22, is the
most difficult since all cross-schema access must be resolved at the service level.
Notice that the database configuration has been changed so that all data access is
done strictly via services and their connected schemas. In this example, Service C
communicates with Service D and not with SchemaD. There is no cross-schema
access; all synonyms created in “Step 2: Assign Tables to Data Domains” on page 156
are removed.
158
|
Chapter 6: Pulling Apart Operational Data
Figure 6-22. Move the cross-schema object access to the services, away from direct crossschema access
When data from other domains is needed, do not reach into their
databases. Instead, access it using the service that owns the data
domain.
Upon completion of this step, the database is in a state of data sovereignty per service,
which occurs when each service owns its own data. Data sovereignty per service is the
nirvana state for a distributed architecture. Like all practices in architecture, it
includes benefits and shortcomings:
Benefits
• Teams can change the database schema without worrying about affecting
changes in other domains.
• Each service can use the database technology and database type best suitable for
their use case.
Shortcomings
• Performance issues occur when services need access to large volumes of data.
• Referential integrity cannot be maintained in the database, resulting in the possi‐
bility of bad data quality.
• All database code (stored procedures, functions) that access tables belonging to
other domains must be moved to the service layer.
Step 4: Move Schemas to Separate Database Servers
Once database teams have created and separated data domains, and have isolated
services so that they access their own data, they can now move the data domains to
separate physical databases. This is often a necessary step because even though
Decomposing Monolithic Data
|
159
services access their own schemas, accessing a single database creates a single archi‐
tecture quantum, as discussed in Chapter 2, which might have adverse effects for
operational characteristics, such as scalability, fault tolerance, and performance.
When moving schemas to separate physical databases, database teams have two
options: backup and restore, or replication. These options are outlined as follows:
Backup and restore
With this option, teams first back up each schema with data domains, then set up
database servers for each data domain. They then restore the schemas, connect
services to schemas in the new database servers, and finally remove schemas
from the original database server. This approach usually requires downtime for
the migration.
Replicate
Using the replicate option, teams first set up database servers for each data
domain. Next they replicate the schemas, switch connections over to the new
database servers, and then remove the schemas from the original database server.
While this approach avoids downtime, it does require more work to set up the
replication and manage increased coordination.
Figure 6-23 shows an example of the replication option, where the database team sets
up multiple database servers so that there is one database server for each data
domain.
Figure 6-23. Replicate schemas (data domains) to their own database servers
160
|
Chapter 6: Pulling Apart Operational Data
Step 5: Switch Over to Independent Database Servers
Once the schemas are fully replicated, the service connections can be switched. The
last step in getting the data domains and services to act as their own independent
deployable units is to remove the connection to the old database servers and remove
the schemas from the old database servers as well. The final state is seen in Figure
6-24.
Figure 6-24. Independent database servers for each data domain
Once the database team has separated the data domains, isolated the database con‐
nections, and finally moved the data domains to their own database servers, they can
optimize the individual database servers for availability and scalability. Teams can
also analyze the data to determine the most appropriate database type to use, intro‐
ducing polyglot database usage within the ecosystem.
Selecting a Database Type
Beginning around 2005, a revolution has occurred in database technologies. Unfortu‐
nately, the number of products that have emerged during this time have created a
problem known as The Paradox of Choice. Having such a large number of products
and choices means having more trade-off decisions to make. Given that each product
is optimized for certain trade-offs, it rests on both software and data architects to pick
the appropriate product with these trade-offs in mind as it relates to their problem
space.
In this section, we introduce star ratings for the various database types, using the fol‐
lowing characteristics in our analysis:
Ease-of-learning curve
This characteristic refers to the ease with which new developers, data architects,
data modelers, operational DBAs, and other users of the databases can learn and
adopt. For example, it’s assumed that most software developers understand SQL,
Selecting a Database Type
|
161
whereas something like Gremlin (a graph query language) may be a niche skill.
The higher the star rating, the easier the learning curve. The lower the star rating,
the harder the learning curve.
Ease of data modeling
This characteristic refers to the ease with which data modelers can represent the
domain in terms of a data model. A higher star rating means data modeling
matches many use cases, and once modeled, is easy to change and adopt.
Scalability/throughput
This characteristic refers to the degree and ease with which a database can scale
to handle increased throughput. Is it easy to scale the database? Can the database
scale horizontally, vertically, or both? A higher star rating means it’s easier to
scale and get higher throughput.
Availability/partition tolerance
This characteristic refers to whether the database supports high availability con‐
figurations (such as replica-sets in MongoDB or tunable consistency in Apache
Cassandra). Does it provide features to handle network partitions? The higher
the star rating, the better the database supports higher availability and/or better
partition tolerance.
Consistency
This characteristic refers to whether the database supports an “always consistent”
paradigm. Does the database support ACID transactions, or does it lean toward
BASE transactions with an eventual consistency model? Does it provide features
to have tunable consistency models for different types of writes? The higher the
star rating, the more consistency the database supports.
Programming language support, product maturity, SQL support, and community
This characteristic refers to which (and how many) programming languages the
database supports, how mature the database is, and the size of the database com‐
munity. Can an organization easily hire people who know how to work with the
database? Higher star ratings means there is better support, the product is
mature, and it’s easy to hire talent.
Read/write priority
This characteristic refers to whether the database prioritizes reads over writes, or
writes over reads, or if it is balanced in its approach. This is not a binary choice—
rather, it’s more of a scale toward which direction the database optimizes.
162
|
Chapter 6: Pulling Apart Operational Data
Relational Databases
Relational databases (also known as an RDBMS) have been the database of choice for
more than three decades. There is significant value in their usage and the stability
they provide, particularly within most business-related applications. These databases
are known for the ubiquitous Structured Query Language (SQL) and the ACID prop‐
erties they provide. The SQL interface they provide makes them a preferred choice
for implementing different read models on top of the same write model. The star
ratings for relational databases appear in Figure 6-25.
Figure 6-25. Relational databases rated for various adoption characteristics
Ease-of-learning curve
Relational databases have been around for many years. They are commonly
taught in schools, and mature documentation and tutorials exist. Therefore, they
are much easier to learn than other database types.
Ease of data modeling
Relational databases allow for flexible data modeling. They allow the modeling of
key-value, document, graph-like structures, and they allow for changes in read
patterns with addition of new indexes. Some models are really difficult to ach‐
ieve, such as graph structures with arbitrary depth. Relational databases organize
data into tables and rows (similar to spreadsheets), something that is natural for
most database modelers.
Selecting a Database Type
|
163
Scalability/throughput
Relational databases are generally vertically scaled using large machines. How‐
ever, setup with replications and automated switchover are complex, requiring
higher coordination and setup.
Availability/partition tolerance
Relational databases favor consistency over availability and partition tolerance,
discussed in “Table Split Technique” on page 254.
Consistency
Relational databases have been dominant for years because of their support for
ACID properties. The ACID features handle many concerns in concurrent sys‐
tems and allow for developing applications without being concerned about
lower-level details of concurrency and how the databases handle them.
Programming language support, product maturity, SQL support, and community
Since relational databases have been around for many years, well-known design,
implementation, and operational patterns can be applied to them, thus making
them easy to adopt, develop, and integrate within an architecture. Many of the
relational databases lack support for reactive stream APIs and similar new con‐
cepts; newer architectural concepts take longer to implement in well-established
relational databases. Numerous programming language interfaces work with
relational databases, and the community of users is large (although splintered
among all the vendors).
Read/write priority
In relational databases, the data model can be designed in such a way that either
reads become more efficient or writes become more efficient. The same database
can handle different types of workloads, allowing for balanced read-write prior‐
ity. For example, not all use cases need ACID properties, especially in large data
and traffic scenarios, or when really flexible schema is desired such as in survey
administration. In these cases, other database types may be a better option.
MySQL, Oracle, Microsoft SQL Server, and PostgreSQL are the most popular rela‐
tional databases and can be run as standalone installations or are available as Data‐
base as a Service on major cloud provider platforms.
164
|
Chapter 6: Pulling Apart Operational Data
Aggregate Orientation
Aggregate orientation is the preference to operate on data that is related and has a
complex data structure. Aggregate is a term originated in Domain-Driven Design:
Tackling Complexity in the Heart of Software by Erik Evans. Think of ticket or cus
tomer with all its dependent tables in the Sysops Squad—they are aggregates. Like all
practices in architecture, aggregate orientation includes benefits and shortcomings:
Benefits
• Enables easy distribution of data in clusters of servers, as the whole aggregate can
be copied over to different servers.
• Improves read and write performance, as it reduces joins in the database.
• Reduces impedance mismatch between the application model and storage model.
Shortcomings
• It’s difficult to arrive at proper aggregates, and changing aggregate boundaries is
hard.
• Analyzing data across aggregates is difficult.
Key-Value Databases
Key-value databases are similar to a hash table data structure, something like tables in
an RDBMS with an ID column as the key and a blob column as the value, which can
consequently store any type of data. Key-value databases are part of a family known
as NoSQL databases. In the book NoSQL Distilled: A Brief Guide to the Emerging
World of Polyglot Persistence (Addison-Wesley), Pramod Sadalage (one of your
authors) and Martin Fowler describe the rise of NoSQL databases and the motiva‐
tions, usages, and trade-offs of using these types of databases, and is a good reference
for further information on this database type.
Key-value databases are easiest to understand among the NoSQL databases. An appli‐
cation client can insert a key and a value, get a value for a known key, or delete a
known key and its value. A key-value database does not know what’s inside the value
part, nor does it care what’s inside, meaning that the database can query using the key
and nothing else.
Unlike relational databases, key-value databases should be picked based on needs.
There are persistent key-value databases like Amazon DynamoDB or Riak KV, non‐
persistent databases like MemcacheDB, and other databases like Redis that can be
configured to be persistent or not. Other relational database constructs like joins,
where, and order by are not supported, but rather the operations get, put, and
delete. The ratings for key-value databases appear in Figure 6-26.
Selecting a Database Type
|
165
Figure 6-26. Key-value databases rated for various adoption characteristics
Ease-of-learning curve
Key-value databases are easy to understand. Since they use “Aggregate Orienta‐
tion” on page 165, it’s important to design the aggregate properly because any
change in the aggregate means rewriting all the data. Moving from relational
databases to any of the NoSQL databases takes practice and unlearning familiar
practices. For example, a developer cannot simply query “Get me all the keys.”
Ease of data modeling
Since key-value databases are aggregate oriented, they can use memory structures
like arrays, maps, or any other type of data, including big blob. The data can be
queried only by key or ID, which means the client should have access to the key
outside of the database. Good examples of a key include session_id, user_id,
and order_id.
Scalability/throughput
Since key-value databases are indexed by key or ID, key lookups are very fast as
there are no joins or order by operations. The value is fetched and returned to
the client, which allows for easier scaling and higher throughput.
Availability/partition tolerance
Since there are many types of key-value databases and each has different proper‐
ties, even the same database can be configured to act in different ways either for
an installation or for each read. For example, in Riak users can use quorum prop‐
erties such as all, one, quorum, and default. When we use one quorum, the
query can return success when any one node responds. When the all quorum
166
|
Chapter 6: Pulling Apart Operational Data
is used, all nodes have to respond for the query to return success. Each query
can tune the partition tolerance and availability. Hence, assuming that all keyvalue stores are the same is a mistake.
Consistency
During each write operation, we can apply configurations that are similar to
applying quorum during read operations; these configurations provide what is
known as tunable consistency. Higher consistency can be achieved by trading off
latency. For a write to be highly consistent, all nodes have to respond, which
reduces partition tolerance. Using a majority quorum is considered a good
trade-off.
Programming language support, product maturity, SQL support, and community
Key-value databases have good programming language support, and many open
source databases have an active community to help learn and understand them.
Since most databases have an HTTP REST API, they are much easier to interface
with.
Read/write priority
Since key-value databases are aggregate oriented, access to data via a key or ID is
geared toward read priority. Key-value databases can be used for session storage,
and can be used to cache user properties and preferences as well.
Sharding in Databases
The concept of partitioning is well-known in relational databases: the table data is
partitioned into sets based on a schema on the same database server. Sharding is simi‐
lar to partitioning, but data resides on different servers or nodes. Nodes collaborate to
figure out where data exists or where data should be stored based on a sharding key.
The word shard means horizontal partition of data in a database.
Document Databases
Documents such as JSON or XML are the basis of document databases. Documents
are human-readable, self-describing, hierarchical tree structures. Document data‐
bases are another type of NoSQL database, whose ratings appear in Figure 6-27.
These databases understand the structure of the data and can index multiple
attributes of the documents, allowing for better query flexibility.
Selecting a Database Type
|
167
Figure 6-27. Document databases rated for various adoption characteristics
Ease-of-learning curve
Document databases are like key-value databases where the value is human read‐
able. This makes learning the database much easier. Enterprises are used to deal‐
ing with documents, such as XML and JSON in different contexts, such as API
payloads and JavaScript frontends.
Ease of data modeling
Just like key-value databases, data modeling involves modeling aggregates such as
orders, tickets, and other domain objects. Document databases are forgiving
when it comes to aggregate design, as the parts of the aggregate are queryable and
can be indexed.
Scalability/throughput
Document databases are aggregate oriented and easy to scale. Complex indexing
reduces scalability, and increased data size leads to a need for partitioning or
sharding. Once sharding is introduced, it increases the complexity and also
forces the selection of a sharding key.
Availability/partition tolerance
Like key-value databases, document databases can be configured for higher avail‐
ability. The setup gets complicated when there are replicated clusters for sharded
collections. The cloud providers are trying to make these setups more usable.
168
|
Chapter 6: Pulling Apart Operational Data
Consistency
Some document databases have started supporting ACID transactions within a
collection, but this may not work in some edge cases. Just like key-value data‐
bases, document databases provide the ability to tune the read and write opera‐
tions using the quorum mechanism.
Programming language support, product maturity, SQL support, and community
Document databases are the most popular of the NoSQL databases, with an
active user community, numerous online learning tutorials, and many program‐
ming language drivers that allow for easier adoption.
Read/write priority
Document databases are aggregate oriented and have secondary indexes to query,
so these databases are favoring read priority.
Schema-less Databases
One common theme in NoSQL databases is duplication of data and schema attribute
names. No two entries have to be the same in terms of schema or attribute names.
This introduces interesting change control dynamics and provides flexibility. The
schema-less nature of the database is powerful, but it’s important to understand that
the data always has a schema even if it’s implicit or defined elsewhere. The application
needs to handle multiple versions of the schema returned by a database. The claim
that NoSQL databases are entirely schema-less is misleading.
Column Family Databases
Column family databases, also known as wide column databases or big table databases,
have rows with varying numbers of columns, where each column is a name-value
pair. With columnar databases, the name is known as a column-key, the value is
known as a column-value, and the primary key of a row is known as a row key. Col‐
umn family databases are another type of NoSQL database that group related data
that is accessed at the same time, and whose ratings appear in Figure 6-28.
Selecting a Database Type
|
169
Figure 6-28. Column family databases rated for various adoption characteristics
Ease-of-learning curve
Column family databases are difficult to understand. Since a collection of namevalue pairs belong to a row, each row can have different name-value pairs. Some
name-value pairs can have a map of columns and are known as super columns.
Understanding how to use these takes practice and time.
Ease of data modeling
Data modeling with column family databases takes some getting used to. Data
needs to be arranged in groups of name-value pairs that have a single row identi‐
fier, and designing this row key takes multiple iterations. Some column family
databases like Apache Cassandra have introduced a SQL-like query language
known as Cassandra Query Language (CQL) that makes data modeling
accessible.
Scalability/throughput
All column family databases are highly scalable and suit use cases where high
write or read throughput is needed. Column family databases scale horizontally
for read and write operations.
Availability/partition tolerance
Column family databases naturally operate in clusters, and when some nodes of
the cluster are down, it is transparent to the client. The default replication factor
is three, which means at least three copies of data are made, improving availabil‐
ity and partition tolerance. Similar to key-value and document databases, column
family databases can tune writes and reads based on quorum needs.
170
|
Chapter 6: Pulling Apart Operational Data
Consistency
Column family databases, like other NoSQL databases, follow the concept of tun‐
able consistency. This means that, based on needs, each operation can decide
how much consistency is desired. For example, in high write scenarios where
some data loss can be tolerated, the write consistency level of ANY could be used,
which means at least one node has accepted the write, while a consistency level of
ALL means all nodes have to accept the write and respond success. Similar consis‐
tency levels can be applied to read operations. It’s a trade-off—higher consistency
levels reduce availability and partition tolerance.
Programming language support, product maturity, SQL support, and community
Column family databases like Cassandra and Scylla have active communities, and
the development of SQL-like interfaces has made the adoption of these databases
easier.
Read/write priority
Column family databases use the concepts of SSTables, commit logs, and memta‐
bles, and since the name-value pairs are populated when data is present, they can
handle sparse data much better than relational databases. They are ideal for high
write-volume scenarios.
All NoSQL databases are designed to understand aggregate orientation. Having
aggregates improves read and write performance, and also allows for higher availabil‐
ity and partition tolerance when the databases are run as a cluster. The notion of CAP
theorem is covered in “Table Split Technique” on page 254 at more length.
Graph Databases
Unlike relational databases, where relations are implied based on references, graph
databases use nodes to store entities and their properties. These nodes are connected
with edges, also known as relationships, which are explicit objects. Nodes are organ‐
ized by relationships and allow for analysis of the connected data by traversing along
specific edges.
Figure 6-29. In graph databases, direction of the edge has significance when querying
Selecting a Database Type
|
171
The edges in graph databases have directional significance. In Figure 6-29, an edge of
type TICKET_CREATED connecting a ticket node with ID 4235143 to a customer node
with ID Neal. We can traverse from the ticket node via the outgoing edge TICKET_CRE
ATED or the customer node via the incoming edge TICKET_CREATED. When the direc‐
tions get mixed up, querying the graph becomes really difficult. The ratings for graph
databases are illustrated in Figure 6-29.
Figure 6-30. Graph databases rated for various adoption characteristics
Ease-of-learning curve
Graph databases have a steep learning curve. Understanding how to use the
nodes, relations, relation type, and properties takes time.
Ease of data modeling
Understanding how to model the domains and convert them into nodes and rela‐
tions is hard. In the beginning, the tendency is to add properties to relations. As
modeling knowledge improves, increased usage of nodes and relations, and con‐
verting some relation properties to nodes with additional relation type takes
place, which improves graph traversal.
Scalability/throughput
Replicated nodes improve read scaling, and throughput can be tuned for read
loads. Since it’s difficult to split or shard graphs, write throughput is constrained
with the type of graph database picked. Traversing the relationships is very fast,
as the indexing and storage is persisted and not calculated at query time.
172
|
Chapter 6: Pulling Apart Operational Data
Availability/partition tolerance
Some of the graph databases that have high partition tolerance and availability
are distributed. Graph database clusters can use nodes that can be promoted as
leaders when current leaders are unavailable.
Consistency
Many graph databases support ACID transactions. Some graph databases, such as
Neo4j, support transactions, so that data is always consistent.
Programming language support, product maturity, SQL support, and community
Graph databases have lots of support in the community. Many algorithms, like
Dijkstra’s algorithm or node similarity, are implemented in the database, reducing
the need to write them from scratch. The language framework known as Gremlin
works across many different databases, helping in the ease of use. Neo4J supports
a query language known as Cypher, allowing developers to easily query the
database.
Read/write priority
In graph databases, data storage is optimized for relationship traversal as
opposed to relational databases, where we have to query the relationships and
derive them at query time. Graph databases are better for read-heavy scenarios.
Graph databases allow the same node to have various types of relationships. In the
Sysops Squad example, a sample graph might look as follows: a knowledge_base was
created_by user sysops_user and knowledge_base used_by sysops_user. Thus, the
relationships created_by and used_by join the same nodes for different relationship
types.
Changing Relationship Types
Changing relationship types is an expensive operation, since each relationship type
has to be re-created. When this happens, both nodes connected by the edge have to be
visited, the new edge created, and the old edge removed. Hence, edge type or relation‐
ship types have to be thought about carefully.
NewSQL Databases
Matthew Aslett first used the term NewSQL to define new databases that aimed to
provide the scalability of NoSQL databases while supporting the features of relational
databases like ACID. NewSQL databases use different types of storage mechanisms,
and all of them support SQL.
NewSQL databases, whose ratings appear in Figure 6-31, improve upon relational
databases by providing automated data partitioning or sharding, allowing for
Selecting a Database Type
|
173
horizontal scaling and improved availability, while at the same time allowing an easy
transition for developers to use the known paradigm of SQL and ACID.
Figure 6-31. New SQL databases rated for various adoption characteristics
Ease-of-learning curve
Since NewSQL databases are just like relational databases (with SQL interface,
added features of horizontal scaling, ACID compliant), the learning curve is
much easier. Some of them are available as only Database as a Service (DBaaS),
which may make learning them more difficult.
Ease of data modeling
Since NewSQL databases are like relational databases, data modeling is familiar
to many and easier to pick up. The extra wrinkle is sharding design, allowing
sharded data placement in geographically different locations.
Scalability/throughput
NewSQL databases are designed to support horizontal scaling for distributed sys‐
tems, allowing for multiple active nodes, unlike relational databases that have
only one active leader, and the rest of the nodes are followers. The multiple active
nodes allow NewSQL databases to be highly scalable and to have better
throughput.
Availability/partition tolerance
Because of the multiple active nodes design, the benefits to availability can be
really high with greater partition tolerance. CockroachDB is a popular NewSQL
database that survives disk, machine, and data center failures.
174
|
Chapter 6: Pulling Apart Operational Data
Consistency
NewSQL databases support strongly consistent ACID transactions. The data is
always consistent, and this allows for relational database users to easily transition
to NewSQL databases.
Programming language support, product maturity, SQL support, and community
There are many open source NewSQL databases, so learning them is accessible.
Some of the databases also support wire-compatible protocols with existing rela‐
tional databases, which allows them to replace relational databases without any
compatibility problems.
Read/write priority
NewSQL databases are used just like relational databases, with added benefits of
indexing and distributing geographically either to improve read performance or
write performance.
Cloud Native Databases
With increased cloud usage, cloud databases such as Snowflake, Amazon Redshift,
Datomic, and Azure CosmosDB have gained in popularity. These databases reduce
operational burden, provide cost transparency, and are an easy way to experiment
since no up-front investments are needed. Ratings for cloud native databases appear
in Figure 6-32.
Figure 6-32. Cloud native databases rated for various adoption characteristics
Selecting a Database Type
|
175
Ease-of-learning curve
Some cloud databases like AWS Redshift are like relational databases and there‐
fore are easier to understand. Databases like Snowflake, which have a SQL inter‐
face but have different storage and compute mechanisms, require some practice.
Datomic is totally different in terms of models and uses immutable atomic facts.
Thus, the learning curve varies with each database offering.
Ease of data modeling
Datomic does not have the concept of tables or the need to define attributes in
advance. It is necessary to define properties of individual attributes, and entities
can have any attribute. Snowflake and Redshift are used more for data warehous‐
ing type workloads. Understanding the type of modeling provided by the data‐
base is critical in selecting the database to use.
Scalability/throughput
Since all these databases are cloud only, scaling them is relatively simple since
resources can be allocated automatically for a price. In these decisions, the tradeoff typically relates to price.
Availability/partition tolerance
Databases in this category (such as Datomic) are highly available when deployed
using Production Topology. They have no single point of failure and are sup‐
ported by extensive caching. Snowflake, for example, replicates its databases
across regions and accounts. Other databases in this category support higher
availability with various options to configure. For example, Redshift runs in a
single availability zone and would need to be run in multiple clusters to support
higher availability.
Consistency
Datomic supports ACID transactions using storage engines to store blocks in
block storage. Other databases, like Snowflake and Redshift, support ACID
transactions.
Programming language support, product maturity, SQL support, and community
Many of these databases are new, and finding experienced help can be difficult.
Experimenting with these databases requires a cloud account, which can create
another barrier. While cloud native databases reduce operational workload on
the operational DBAs, they do have a higher learning curve for developers.
Datomic uses Clojure in all its examples, and stored procedures are written with
Clojure, so not knowing Clojure maybe a barrier to usage.
Read/write priority
These databases can be used for both read-heavy or write-heavy loads. Snowflake
and Redshift are geared more toward data warehouse type workloads, lending
176
|
Chapter 6: Pulling Apart Operational Data
them toward read priority, while Datomic can support both type of loads with
different indexes such as EAVT (Entity, Attribute, Value, then Transaction) first.
Time-Series Databases
Given the trends, we see increased usage of IoT, microservices, self-driving cars, and
observability, all of which have driven a phenomenal increase in time-series analytics.
This trend has given rise to databases optimized for storing sequences of data points
collected during a time window, enabling users to track changes over any duration of
time. The ratings for this database type appear in Figure 6-33.
Figure 6-33. Time-series databases rated for various adoption characteristics
Ease-of-learning curve
Understanding time-series data is often easy—every data point is attached to a
timestamp, and data is almost always inserted and never updated or deleted.
Understanding append-only operations takes some unlearning from other data‐
base usage, where errors in the data can be corrected with an update. InfluxDB,
Kx, and TimeScale are some of the popular time-series databases.
Ease of data modeling
The underlying concept with time-series databases is to analyze changes in data
over time. For example, with the Sysops Squad example, changes done to a ticket
object can be stored in a time-series database, where the timestamp of change
and ticket_id are tagged. It’s considered bad practice to add more than one
piece of information in one tag. For example, ticket_status=Open,
ticket_id=374737 is better than ticket_info=Open.374737.
Selecting a Database Type
|
177
Scalability/throughput
Timescale is based on PostgreSQL and allows for standard scaling and through‐
put improvement patterns. Running InfluxDB in cluster mode by using meta
nodes that manage metadata and data nodes that store actual data provides scal‐
ing and throughput improvements.
Availability/partition tolerance
Some databases like InfluxDB have better availability and partition tolerance
options with configurations for meta and data nodes, along with replication
factors.
Consistency
Time-series databases that use relational databases as their storage engine get
ACID properties for consistency, while other databases can tune consistency
using consistency-level of any, one, or quorum. Higher consistency-level con‐
figuration generally results in higher consistency and lower availability, so it’s a
trade-off that needs to be considered.
Programming language support, product maturity, SQL support, and community
Time-series databases have become popular lately, and there are many resources
to learn from. Some of these databases, such as InfluxDB, provide a SQL-like
query language known as InfluxQL.
Read/write priority
Time-series databases are append only and tend to be better suited for readheavy workloads.
When using time-series databases, the database automatically attaches a timestamp to
every datum creation, and the data contains tags or attributes of information. The
data is queried based on some fact between specific time windows. Therefore, timeseries databases are not general-purpose databases.
In summary of all the database types discussed in this section, Table 6-6 shows some
popular database products for the database type.
Table 6-6. Summary of database types and products in the database type
Database type
Products
Relational
Key-value
Document
Column family
Graph
NewSQL
PostgreSQL, Oracle, Microsoft SQL
Riak KV, Amazon DynamoDB, Redis
MongoDB, Couchbase, AWS DocumentDB
Cassandra, Scylla, Amazon SimpleDB
Neo4j, Infinite Graph, Tiger Graph
VoltDB, ClustrixDB, SimpleStore (aka MemSQL)
178
|
Chapter 6: Pulling Apart Operational Data
Database type
Products
Cloud native
Time-series
Snowflake, Datomic, Redshift
InfluxDB, kdb+, Amazon Timestream
Sysops Squad Saga: Polyglot Databases
Thursday, December 16, 16:05
Now that the team had formed data domains from the monolithic Sysops Squad database, Devon noticed that the Survey data domain would be a great candidate for
migrating from a traditional relational database to a document database using JSON.
However, Dana, the head of data architecture, didn’t agree and wanted to keep the
tables as relational.
“I simply don’t agree,” said Dana. “The survey tables have always worked in the past as
relational tables, so I see no reason to change things around. "
“Actually,” said Skyler, “if you had originally talked with us about this when the system was first being
developed, you would understand that from a user interface perspective, it’s really hard to deal with
relational data for something like a customer survey. So I disagree. It may work out good for you, but
from a user interface development standpoint, dealing with relational data for the survey stuff has
been a major pain point.”
“See, so there you are,” said Devon. “This is why we need to change it to a document database.”
“You seem to forget that as the data architect for this company, I am the one who has ultimate
responsibility for all these different databases. You can’t just start adding different database types to
the system,” said Dana.
“But it would be a much better solution,” said Devon.
“Sorry, but I’m not going to cause a disruptor on the database teams just so Skyler can have an easier
job maintaining the user interface. Things don’t work that way.”
“Wait,” said Skyler, “didn’t we all agree that part of the problem of the current monolithic Sysops
Squad application was that the development teams didn’t work close enough with the database
teams?”
“Yes,” said Dana.
“Well then,” said Skyler, “let’s do that. Let’s work together to figure this out.”
“OK,” said Dana, “but what I’m going to need from you and Devon is a good solid justification for
introducing another type of database into the mix.”
“You got it,” said Devon. “We’ll start working on that right away.”
Sysops Squad Saga: Polyglot Databases
|
179
Devon and Skyler knew that a document database would be a much better solution for the customer survey data, but they weren’t sure how to build the right justifications for Dana to agree to
migrate the data. Skyler suggested that they meet with Addison to get some help, since both
agreed that this was somewhat an architectural concern. Addison agreed to help, and set up a
meeting with Parker (the Sysops Squad product owner) to validate whether there was any business
justification to migrating the customer survey tables to a document database.
“Thanks for meeting with us, Parker,” said Addison. “As I mentioned to you before, we are thinking of
changing the way the customer survey data is stored, and have a few questions for you.”
“Well,” said Parker, “that was one of the reasons why I agreed to this meeting. You see, the customer
survey part of the system has been a major pain point for the marketing department, as well as for
me.”
“Huh?” asked Skyler. “What do you mean?”
“How long does it take you to apply even the smallest of change requests to the customer surveys?”
asked Parker.
“Well,” said Devon, “it’s not too bad from the database side. I mean, it’s a matter of adding a new
column for a new question or changing the answer type.”
“Hold on,” said Skyler. “Sorry, but for me it’s a major change, even when you add an additional question. You have no idea how hard it is to query all of that relational data and render a customer survey
in the user interface. So, my answer is, a very long time.”
“Listen,” said Parker. “We on the business side of things get very frustrated ourselves when even the
simplest of changes take you literally days to do. It’s simply not acceptable.”
“I think I can help here,” said Addison. “So Parker, what you’re saying is that the customer survey
changes frequently, and it is taking too long to make the changes?”
“Correct,” said Parker. “The marketing department not only wants better flexibility in the customer
surveys, but better response from the IT department as well. Many times they don’t place change
requests because they know it will just end in frustration and additional cost they didn’t plan for.”
“What if I were to tell you that the lack of flexibility and responsiveness to change requests has
everything to do with the technology used to store customer surveys, and that by changing the way
we store data, we could significantly improve flexibility as well as response time for change
requests?” asked Addison.
“Then I would be the happiest person on Earth, as would the marketing department,” said Parker.
“Devon and Skyler, I think we have our business justification,” said Addison.
With the business justification established, Devon, Skyler, and Addison convinced Dana to use a
document database. Now the team had to figure out the optimal structure for the customer survey
data. The existing relational database tables are illustrated in Figure 6-34. Each customer survey
180
|
Chapter 6: Pulling Apart Operational Data
consisted of two primary tables—a Survey table and a Question table, with a one-to-many relationship between the two tables.
Figure 6-34. Tables and relationships in the sysops survey data domain
An example of the data contained in each table is shown Figure 6-35, where the Question table contains the question, the answer options, and the data type for the answer.
Figure 6-35. Relational data in tables for survey and question in the survey data domain
“So, essentially we have two options for modeling the survey questions in a document database,”
said Devon. “A single aggregate document or one that is split.”
“How to we know which one to use?” asked Skyler, happy that the development teams were now
finally working with the database teams to arrive at a unified solution.
“I know,” said Addison, “let’s model both so we can visually see the trade-offs with each approach.”
Devon showed the team that with the single aggregate option, as shown in Figure 6-36, with the
corresponding source code listing in Example 6-3, both the survey data and all related question data
were stored as one document. Therefore, the entire customer survey could be retrieved from the
database by using a single get operation, making it easy for Skyler and others on the development
team to work with the data.
Sysops Squad Saga: Polyglot Databases
|
181
Figure 6-36. Survey model with single aggregate
Example 6-3. JSON document for single aggregate design with children embedded
# Survey aggregate with embedded questions
{
"survey_id": "19999",
"created_date": "Dec 28 2021",
"description": "Survey to gauge customer...",
"questions": [
{
"question_id": "50001",
"question": "Rate the expert",
"answer_type": "Option",
"answer_options": "1,2,3,4,5",
"order": "2"
},
{
"question_id": "50000",
"question": "Did the expert fix the problem?",
"answer_type": "Boolean",
"answer_options": "Yes,No",
"order": "1"
}
]
}
182
| Chapter 6: Pulling Apart Operational Data
“I really like that approach,” said Skyler. “Essentially, I wouldn’t have to worry so much about aggregating things myself in the user interface, meaning I could simply render the document I retrieve on
the web page.”
“Yeah,” said Devon, “but it would require additional work on the database side as questions would be
replicated in each survey document. You know, the whole reuse argument. Here, let me show you
the other approach.”
Skyler explained that another way to think about aggregates was to split the survey and question
model so that the questions could be operated on in an independent fashion, as shown in Figure
6-37, with the corresponding source code listing in Example 6-4. This would allow the same question to be used in multiple surveys, but would be harder to render and retrieve than the single
aggregate.
Figure 6-37. Survey model with multiple aggregates with references
Example 6-4. JSON document with aggregates split and parent document showing
references to children
# Survey aggregate with references to Questions
{
"survey_id": "19999",
"created_date": "Dec 28",
"description": "Survey to gauge customer...",
"questions": [
{"question_id": "50001", "order": "2"},
{"question_id": "50000", "order": "1"}
]
Sysops Squad Saga: Polyglot Databases
|
183
}
# Question aggregate
{
"question_id": "50001",
"question": "Rate the expert",
"answer_type": "Option",
"answer_options": "1,2,3,4,5"
}
{
"question_id": "50000",
"question": "Did the expert fix the problem?",
"answer_type": "Boolean",
"answer_options": "Yes,No"
}
Because most of the complexity and change issues were in the user interface, Skyler liked the single
aggregate model better. Devon liked the multiple aggregate to avoid duplication of question data in
each survey. However, Addison pointed out that there were only five survey types (one for each
product category), and that most of the changes involved adding or removing questions. The team
discussed the trade-offs, and all agreed that they were willing to trade off some duplication of question data for the ease of changes and rendering on the user interface side. Because of the difficulty
of this decision and the structural nature of changing the data, Addison created an ADR to record
the justifications of this decision:
ADR: Use of Document Database for Customer Survey
Context
Customers receive a survey after the work has been completed by the customer, which is
rendered on a web page for the customer to fill out and submit. The customer receives
one of five survey types based on the type of electronic product fixed or installed. The
survey is currently stored in a relational database, but the team wants to migrate the survey to a document database using JSON.
Decision
We will use a document database for the customer survey.
The Marketing Department requires more flexibility and timeliness for changes to the customer surveys. Moving to a document database would not only provide better flexibility,
but also better timeliness for changes needed to the customer surveys.
Using a document database would simplify the customer survey user interface and better
facilitate changes to the surveys.
Consequences
Since we will be using a single aggregate, multiple documents would need to be changed
when a common survey question is updated, added, or removed.
Survey functionality will need to be shut down during the data migration from the relational database to the document database.
184
| Chapter 6: Pulling Apart Operational Data
CHAPTER 7
Service Granularity
Thursday, October 14, 13:33
As the migration effort got underway, both Addison and Austen started getting overwhelmed with all of the decisions involved with breaking apart the domain services previously identified. The development team also had its own opinions, which made
decision making for service granularity even more difficult.
“I’m still not sure what to do with the core ticketing functionality,” said Addison. “I can’t
decide whether ticket creation, completion, expert assignment, and expert routing
should be one, two, three, or even four services. Taylen is insisting on making everything
fine-grained, but I’m not sure that’s the right approach.”
“Me neither,” said Austen. “And I’ve got my own issues trying to figure out if the customer registration, profile management, and billing functionality should even be broken apart. And on top of all
that, I’ve got another game this evening.”
“You’ve always got a game to go to,” said Addison. “Speaking of customer functionality, did you ever
figure out if the customer login functionality is going to be a separate service?”
“No,” said Austen, “I’m still working on that as well. Skyler says it should be separate, but won’t give
me a reason other than to say it’s separate functionality.”
“This is hard stuff,” said Addison. “Do you think Logan can shed any light on this?”
“Good idea,” said Austen, “This seat-of-the-pants analysis is really slowing things down.”
Addison and Austen invited Taylen, the Sysops Squad tech lead, to the meeting with Logan so that
all of them could be on the same page with regard to the service granularity issues they were facing.
185
“I’m telling you,” said Taylen, “we need to break up the domain services into smaller services. They are
simply too coarse-grained for microservices. From what I remember, micro means small. We are, after
all, moving to microservices. What Addison and Austen are suggesting simply doesn’t fit with the
microservices model.”
“Not every portion of an application has to be microservices,” said Logan. “That’s one of the biggest
pitfalls of the microservices architecture style.”
“If that’s the case, then how do you determine what services should and shouldn’t be broken apart?”
asked Taylen.
“Let me ask you something, Taylen,” said Logan. “What is your reason for wanting to make all of the
services so small?”
“Single-responsibility principle,” answered Taylen. “Look it up. That’s what microservices is based on.”
“I know what the single-responsibility principle is,” said Logan. “And I also know how subjective it can
be. Let’s take our customer notification service as an example. We can notify our customers through
SMS, email, and we even send out postal letters. So tell me everyone, one service or three services?”
“Three,” immediately answered Taylen. “Each notification method is its own thing. That’s what microservices is all about.”
“One,” answered Addison. “Notification itself is clearly a single responsibility.”
“I’m not sure,” answered Austen. “I can see it both ways. Should we just toss a coin?”
“This is exactly why we need help,” sighed Addison.
“The key to getting service granularity right,” said Logan, “is to remove opinion and gut feeling, and
use granularity disintegrators and integrators to objectively analyze the trade-offs and form solid
justifications for whether or not to break apart a service.”
“What are granularity disintegrators and integrators?” asked Austen.
“Let me show you,” said Logan.
Architects and developers frequently confuse the terms modularity and granularity,
and in some cases even treat them to mean the same thing. Consider the following
dictionary definitions of each of these terms:
Modularity
Constructed with standardized units or dimensions for flexibility and variety in
use.
Granularity
Consisting of or appearing to consist of one of numerous particles forming a
larger unit.
186
|
Chapter 7: Service Granularity
It’s no wonder so much confusion exists between these terms! Although the terms
have similar dictionary definitions, we want to distinguish between them because
they mean different things within the context of software architecture. In our usage,
modularity concerns breaking up systems into separate parts (see Chapter 3), whereas
granularity deals with the size of those separate parts. Interestingly enough, most
issues and challenges within distributed systems are typically not related to modular‐
ity, but rather granularity.
Determining the right level of granularity—the size of a service—is one of the many
hard parts of software architecture that architects and development teams continually
struggle with. Granularity is not defined by the number of classes or lines of code in a
service, but rather what the service does—hence why it is so hard to get service gran‐
ularity right.
Architects can leverage metrics to monitor and measure various aspects of a service
to determine the appropriate level of service granularity. One such metric used to
objectively measure the size of a service is to calculate the number of statements in a
service. Every developer has a different coding style and technique, which is why the
number of classes and number of lines of code are poor metrics to use to measure
granularity. The number of statements, on the other hand, at least allows an architect
or development team to objectively measure what the service is doing. Recall from
Chapter 4 that a statement is a single complete action performed in the source code,
usually terminated by a special character (such as a semicolon in languages such as
Java, C, C++, C#, Go, JavaScript; or a newline in languages such as F#, Python, and
Ruby).
Another metric to determine service granularity is to measure and track the number
of public interfaces or operations exposed by a service. Granted, while there is still a
bit of subjectiveness and variability with these two metrics, it’s the closest thing we’ve
come up with so far to objectively measure and assess service granularity.
Two opposing forces for service granularity are granularity disintegrators and granu‐
larity integrators. These opposing forces are illustrated in Figure 7-1. Granularity dis‐
integrators address the question “When should I consider breaking apart a service
into smaller parts?”, whereas Granularity integrators address the question “When
should I consider putting services back together?” One common mistake many devel‐
opment teams make is focusing too much on granularity disintegrators while ignor‐
ing granularity integrators. The secret of arriving at the appropriate level of
granularity for a service is achieving an equilibrium between these two opposing
forces.
Service Granularity
|
187
Figure 7-1. Service granularity depends on a balance of disintegrators and integrators
Granularity Disintegrators
Granularity disintegrators provide guidance and justification for when to break a ser‐
vice into smaller pieces. While the justification for breaking up a service may involve
only a single driver, in most cases the justification will be based on multiple drivers.
The six main drivers for granularity disintegration are as follows:
Service scope and function
Is the service doing too many unrelated things?
Code volatility
Are changes isolated to only one part of the service?
Scalability and throughput
Do parts of the service need to scale differently?
Fault tolerance
Are there errors that cause critical functions to fail within the service?
Security
Do some parts of the service need higher security levels than others?
Extensibility
Is the service always expanding to add new contexts?
188
|
Chapter 7: Service Granularity
The following sections detail each of these granularity disintegration drivers.
Service Scope and Function
The service scope and function is the first and most common driver for breaking up a
single service into smaller ones, particularly with regard to microservices. There are
two dimensions to consider when analyzing the service scope and function. The first
dimension is cohesion: the degree and manner to which the operations of a particular
service interrelate. The second dimension is the overall size of a component, meas‐
ured usually in terms of the total number of statements summed from the classes that
make up that service, the number of public entrypoints into the service, or both.
Consider a typical Notification Service that does three things: notifies a customer
through SMS (Short Message Service), email, or a printed postal letter that is mailed
to the customer. Although it is very tempting to break this service into three separate
single-purpose services (one for SMS, one for email, and one for postal letters) as
illustrated in Figure 7-2, this alone is not enough to justify breaking the service apart
because it already has relatively strong cohesion—all of these functions relate to one
thing, notifying the customer. Because “single purpose” is left for individual opinion
and interpretation, it is difficult to know whether to break apart this service or not.
Figure 7-2. A service with relatively strong cohesion is not a good candidate for disinte‐
gration based on functionality alone
Granularity Disintegrators
|
189
Now consider a single service that manages the customer profile information, cus‐
tomer preferences, and also customer comments made on the website. Unlike the
previous Notification Service example, this service has relatively weak cohesion
because these three functions relate to a broader scope—customer. This service is
possibly doing too much, and hence should probably be broken into three separate
services, as illustrated in Figure 7-3.
Figure 7-3. A service with relatively weak cohesion is a good candidate for disintegration
This granularity disintegrator is related to the single-responsibility principle coined
by Robert C. Martin as part of his SOLID principles, which states, “every class should
have responsibility over a single part of that program’s functionality, which it should
encapsulate. All of that module, class or function’s services should be narrowly
aligned with that responsibility.” While the single-responsibility principle was origi‐
nally scoped within the context of classes, in later years it has expanded to include
components and services.
190
| Chapter 7: Service Granularity
Within the microservices architecture style, a microservice is defined as a singlepurpose, separately deployed unit of software that does one thing really well. No won‐
der developers are so tempted to make services as small as possible without
considering why they are doing so! The subjectiveness related to what is and isn’t a
single responsibility is where most developers get into trouble with regard to service
granularity. While there are some metrics (such as LCOM) to measure cohesion, it is
nevertheless highly subjective when it comes to services—is notifying the customer
one single thing, or is notifying via email one single thing? For this reason, it is vital
to understand other granularity disintegrators to determine the appropriate level of
granularity.
Code Volatility
Code volatility--the rate at which the source code changes—is another good driver for
breaking a service into smaller ones. This is also known as volatility-based decomposi‐
tion. Objectively measuring the frequency of code changes in a service (easily done
through standard facilities in any source code version-control system) can sometimes
lead to a good justification for breaking apart a service. Consider the notification ser‐
vice example again from the prior section. Service scope (cohesion) alone was not
enough to justify breaking the service apart. However, by applying change metrics,
relevant information is revealed about the service:
• SMS notification functionality rate of change: every six months (avg)
• Email notification functionality rate of change: every six months (avg)
• Postal letter notification functionality rate of change: weekly (avg)
Notice that the postal letter functionality changes weekly (on average), whereas the
SMS and email functionality rarely changes. As a single service, any change to the
postal letter code would require the developer to test and redeploy the entire service,
including SMS and email functionality. Depending on the deployment environment,
this also might mean SMS and email functionality would not be available when the
postal letter changes are deployed. Thus, as a single service, testing scope is increased
and deployment risk is high. However, by breaking this service into two separate
services (Electronic Notification and Postal Letter Notification), as illustrated in Fig‐
ure 7-4, frequent changes are now isolated into a single, smaller service. This in turn
means that the testing scope is significantly reduced, deployment risk is lower, and
SMS and email functionality is not disrupted during a deployment of postal letter
changes.
Granularity Disintegrators
|
191
Figure 7-4. An area of high code change in a service is a good candidate for
disintegration
Scalability and Throughput
Another driver for breaking up a service into separate smaller ones is scalability and
throughput. The scalability demands of different functions of a service can be objec‐
tively measured to qualify whether a service should be broken apart. Consider once
again the Notification Service example, where a single service notifies customers
through SMS, email, and printed postal letter. Measuring the scalability demands of
this single service reveals the following information:
• SMS notification: 220,000/minute
• Email notification: 500/minute
• Postal letter notification: 1/minute
Notice the extreme variation between sending out SMS notifications and postal letter
notifications. As a single service, email and postal letter functionality must unneces‐
sarily scale to meet the demands of SMS notifications, impacting cost and also elastic‐
ity in terms of mean time to startup (MTTS). Breaking the Notification Service into
three separate services (SMS, Email, and Letter), as illustrated in Figure 7-5, allows
each of these services to scale independently to meet their varying demands of
throughput.
192
|
Chapter 7: Service Granularity
Figure 7-5. Differing scalability and throughput needs is a good disintegration driver
Fault Tolerance
Fault tolerance describes the ability of an application or functionality within a particu‐
lar domain to continue to operate, even though a fatal crash occurs (such as an outof-memory condition). Fault Tolerance is another good driver for granularity
disintegration.
Consider the same consolidated Notification Service example that notifies customers
through SMS, email, and postal letter (Figure 7-6). If the email functionality contin‐
ues to have problems with out-of-memory conditions and fatally crashes, the entire
service comes down, including SMS and postal letter processing.
Separating this single consolidated Notification Service into three separate services
provides a level of fault tolerance for the domain of customer notification. Now, a
fatal error in the functionality of the email service doesn’t impact SMS or postal
letters.
Notice in this example that the Notification Service is split into three separate services
(SMS, Email, and Postal Letter), even though email functionality is the only issue
with regard to frequent crashes (the other two are very stable). Since email function‐
ality is the only issue, why not combine the SMS and postal letter functionality into a
single service?
Consider the code volatility example from the prior section. In this case Postal Letter
changes constantly, whereas the other two (SMS and Email) do not. Splitting this
Granularity Disintegrators
|
193
service into only two services made sense because Postal Letter was the offending
functionality, but Email and SMS are related—they both have to do with electronically
notifying the customer. Now consider the fault-tolerance example. What do SMS
notification and Postal Letter notification have in common other than a notification
means to the customer? What would be an appropriate self-descriptive name of that
combined service?
Figure 7-6. Fault tolerance and service availability are good disintegration drivers
Moving the email functionality to a separate service disrupts the overall domain cohe‐
sion because the resulting cohesion between SMS and postal letter functionality is
weak. Consider what the likely service names would be: Email Service and…Other
Notification Service? Email Service and…SMS-Letter Notification Service? Email Ser‐
vice and…Non-Email Service? This naming problem relates back to the service scope
and function granularity disintegrator—if a service is too hard to name because it’s
doing multiple things, then consider breaking apart the service. The following disin‐
tegrations help in visualizing this important point:
• Notification Service → Email Service, Other Notification Service (poor name)
• Notification Service → Email Service, Non-Email Service (poor name)
• Notification Service → Email Service, SMS-Letter Service (poor name)
• Notification Service → Email Service, SMS Service, Letter Service (good names)
194
|
Chapter 7: Service Granularity
In this example, only the last disintegration makes sense, particularly considering the
addition of another social media notification—where would that go? Whenever
breaking apart a service, regardless of the disintegration driver, always check to see if
strong cohesion can be formed with the “leftover” functionality.
Security
A common pitfall when securing sensitive data is to think only in terms of the storage
of that data. For example, securing PCI (Payment Card Industry) data from non-PCI
data might be addressed through separate schemas or databases residing in different
secure regions. What is sometimes missing from this practice, however, is also secur‐
ing how that data is accessed.
Consider the example illustrated in Figure 7-7 that describes a Customer Profile Ser‐
vice containing two main functions: customer profile maintenance for adding, chang‐
ing, or deleting basic profile information (name, address, and so on); and customer
credit card maintenance for adding, removing, and updating credit card information.
Figure 7-7. Security and data access are good disintegration drivers
While the credit card data may be protected, access to that data is at risk because the
credit card functionality is joined together with the basic customer profile functional‐
ity. Although the API entry points into the consolidated customer profile service may
differ, nevertheless there is risk that someone entering into the service to retrieve the
customer name might also have access to credit card functionality. By breaking this
service into two separate services, access to the functionality used to maintain credit
Granularity Disintegrators
|
195
card information can be made more secure because the set of credit card operations is
going into only a single-purpose service.
Extensibility
Another primary driver for granularity disintegration is_ extensibility_—the ability
to add additional functionality as the service context grows. Consider a payment ser‐
vice that manages payments and refunds through multiple payment methods, includ‐
ing credit cards, gift cards, and PayPal transactions. Suppose the company wants to
start supporting other managed payment methods, such as reward points, store credit
from returns; and other third-party payment services, such as ApplePay, SamsungPay,
and so on. How easy is it to extend the payment service to add these additional pay‐
ment methods?
These additional payment methods could certainly be added to a single consolidated
payment service. However, every time a new payment method is added, the entire
payment service would need to be tested (including other payment types), and the
functionality for all other payment methods unnecessarily redeployed into produc‐
tion. Thus, with the single consolidated payment service, the testing scope is
increased and deployment risk is higher, making it more difficult to add additional
payment types.
Now consider breaking up the existing consolidated service into three separate serv‐
ices (Credit Card Processing, Gift Card Processing, and PayPal Transaction Process‐
ing), as illustrated in Figure 7-8.
Now that the single payment service is broken into separate services by payment
methods, adding another payment method (such as reward points) is only a matter of
developing, testing, and deploying a single service separate from the others. As a
result, development is faster, testing scope is reduced, and deployment risk is lower.
Our advice is to apply this driver only if it is known ahead of time that additional
consolidated contextual functionality is planned, desired, or part of the normal
domain. For example, with notification, it is doubtful the means of notification would
continually expand beyond the basic notification means (SMS, email, or letter). How‐
ever, with payment processing, it is highly likely that additional payment types would
be added in the future, and therefore separate services for each payment type would
be warranted. Since it is often difficult to sometimes “guess” whether (and when) con‐
textual functionality might expand (such as additional payment methods), our advice
is to wait on this driver as a primary means of justifying a granularly disintegration
until a pattern can be established or confirmation of continued extensibility can be
confirmed.
196
|
Chapter 7: Service Granularity
Figure 7-8. Planned extensibility is a good disintegration driver
Granularity Integrators
Whereas granularity disintegrators provide guidance and justification for when to
break a service into smaller pieces, granularity integrators work in the opposite way—
they provide guidance and justification for putting services back together (or not
breaking apart a service in the first place). Analyzing the trade-offs between disinte‐
gration drivers and integration drivers is the secret to getting service granularity
right. The four main drivers for granularity integration are as follows:
Database transactions
Is an ACID transaction required between separate services?
Workflow and choreography
Do services need to talk to one another? Shared code: Do services need to share
codeamong one another? Database relationships: Although a service can be bro‐
ken apart, can the data it uses be broken apart as well?
The following sections detail each of these granularity integration drivers.
Granularity Integrators
|
197
Database Transactions
Most monolithic systems and course-grained domain services using relational data‐
bases rely on single-unit-of-work database transactions to maintain data integrity and
consistency; see “Distributed Transactions” on page 263 for the details of ACID
(database) transactions and how they differ from BASE (distributed) transactions. To
understand how database transactions impact service granularity, consider the situa‐
tion illustrated in Figure 7-9 where customer functionality has been split into a Cus‐
tomer Profile Service that maintains customer profile information and a Password
Service that maintains password and other security-related information and
functionality.
Notice that having two separate services provides a good level of security access con‐
trol to password information since access is at a service level rather than at a request
level. Access to operations such as changing a password, resetting a password, and
accessing a customer’s password for sign-in can all be restricted to a single service
(and hence the access can be restricted to that single service). However, while this
may be a good disintegration driver, consider the operation of registering a new cus‐
tomer, as illustrated in Figure 7-10.
Figure 7-9. Separate services with atomic operations have better security access control
198
|
Chapter 7: Service Granularity
Figure 7-10. Separate services with combined operations do not support database
(ACID) transactions
When registering a new customer, both profile and encrypted password information
is passed into the Profile Service from a user interface screen. The Profile Service
inserts the profile information into its corresponding database table, commits that
work, and then passes the encrypted password information to the Password Service,
which in turn inserts the password information into its corresponding database table
and commits its own work.
While separating the services provides better security access control to the password
information, the trade-off is that there is no ACID transaction for actions such as reg‐
istering a new customer or unsubscribing (deleting) a customer from the system. If
the password service fails during either of these operations, data is left in an inconsis‐
tent state, resulting in complex error handling (which is also error prone) to reverse
the original profile insert or take other corrective action (see “Transactional Saga Pat‐
terns” on page 324 for the details of eventual consistency and error handling within
distributed transactions). Thus, if having a single-unit-of-work ACID transaction is
required from a business perspective, these services should be consolidated into a sin‐
gle service, as illustrated in Figure 7-11.
Granularity Integrators
|
199
Figure 7-11. A single service supports database (ACID) transactions
Workflow and Choreography
Another common granularity integrator is workflow and choreography--services talk‐
ing to one another (also sometimes referred to as interservice communication or eastwest communications). Communication between services is fairly common and in
many cases necessary in highly distributed architectures like microservices. However,
as services move toward a finer level of granularity based on the disintegration factors
outlined in the previous section, service communication can increase to a point
where negative impacts start to occur.
Issues with overall fault tolerance is the first impact of too much synchronous interservice communication. Consider the diagram in Figure 7-12: Service A communi‐
cates with services B and C, Service B communicates with Service C, Service D
communicates with Service E, and finally Service E communicates with Service C. In
this case, if Service C goes down, all other services become nonoperational because of
a transitive dependency with Service C, creating an issue with overall fault tolerance,
availability, and reliability.
200
| Chapter 7: Service Granularity
Figure 7-12. Too much workflow impacts fault tolerance
Interestingly enough, fault tolerance is one of the granularity disintegration drivers
from the previous section—yet when those services need to talk to one another, noth‐
ing is really gained from a fault-tolerance perspective. When breaking apart services,
always check to see if the functionalities are tightly coupled and dependent on one
another. If it is, then overall fault tolerance from a business request standpoint won’t
be achieved, and it might be best to consider keeping the services together.
Overall performance and responsiveness is another driver for granularity integration
(putting services back together). Consider the scenario in Figure 7-13: a large cus‐
tomer service is split into five separate services (services A through E). While each of
these services has its own collection of cohesive atomic requests, retrieving all of the
customer information collectively from a single API request into a single user inter‐
face screen involves five separate hops when using choreography (see Chapter 11 for
an alternative solution to this problem using orchestration). Assuming 300 ms in net‐
work and security latency per request, this single request would incur an additional
1500 ms just in latency alone! Consolidating all of these services into a single service
would remove the latency, therefore increasing overall performance and
responsiveness.
Granularity Integrators
|
201
Figure 7-13. Too much workflow impacts overall performance and responsiveness
In terms of overall performance, the trade-off for this integration driver is balancing
the need to break apart a service with the corresponding performance loss if those
services need to communicate with one another. A good rule of thumb is to take into
consideration the number of requests that require multiple services to communicate
with one another, also taking into account the criticality of those requests requiring
interservice communication. For example, if 30% of the requests require a workflow
between services to complete the request and 70% are purely atomic (dedicated to
only one service without the need for any additional communication), then it might
be OK to keep the services separate. However, if the percentages are reversed, then
consider putting them back together again. This assumes, of course, that overall per‐
formance matters. There’s more leeway in the case of backend functionality where an
end user isn’t waiting for the request to complete.
The other performance consideration is with regard to the criticality of the request
requiring workflow. Consider the previous example, where 30% of the requests
require a workflow between services to complete the request, and 70% are purely
atomic. If a critical request that requires extremely fast response time is part of that
30%, then it might be wise to put the services back together, even though 70% of the
requests are purely atomic.
Overall reliability and data integrity are also impacted with increased service commu‐
nication. Consider the example in Figure 7-14: customer information is separated
into five separate customer services. In this case, adding a new customer to the sys‐
tem involves the coordination of all five customer services. However, as explained in
a previous section, each of these services has its own database transaction. Notice in
202
| Chapter 7: Service Granularity
Figure 7-14 that services A, B, and C have all committed part of the customer data, but
Service D fails.
Figure 7-14. Too much workflow impacts reliability and data integrity
This creates a data consistency and data integrity issue because part of the customer
data has already been committed, and may have already been acted upon through a
retrieval of that information from another process or even a message sent out from
one of those services broadcasting an action based on that data. In either case, that
data would either have to be rolled back through compensating transactions or
marked with a specific state to know where the transaction left off in order to restart
it. This is very messy situation, one we describe in detail in “Transactional Saga Pat‐
terns” on page 324. If data integrity and data consistency are important or critical to
an operation, it might be wise to consider putting those services back together.
Shared Code
Shared source code is a common (and necessary) practice in software development.
Functions like logging, security, utilities, formatters, converters, extractors, and so on
are all good examples of shared code. However, things can get complicated when
dealing with shared code in a distributed architecture and can sometimes influence
service granularity.
Shared code is often contained in a shared library, such as a JAR file in the Java Eco‐
system, a GEM in the Ruby environment, or a DLL in the .NET environment, and is
typically bound to a service at compile time. While we dive into code reuse patterns
in detail in Chapter 8, here we illustrate only how shared code can sometimes
Granularity Integrators
|
203
influence service granularity and can become a granularity integrator (putting serv‐
ices back together).
Consider the set of five services shown in Figure 7-15. While there may have been a
good disintegrator driver for breaking apart these services, they all share a common
codebase of domain functionality (as opposed to common utilities or infrastructure
functionality). If a change occurs in the shared library, this would eventually necessi‐
tate a change in the corresponding services using that shared library. We say eventu‐
ally because versioning can sometimes be used with shared libraries to provide agility
and backward compatibility (see Chapter 8). As such, all of these separately deployed
services would have to be changed, tested, and deployed together. In these cases, it
might be wise to consolidate these five services into a single service to avoid multiple
deployments, as well as having the service functionality be out of sync based on the
use of different versions of a library.
Figure 7-15. A change in shared code requires a coordinated change to all services
Not all uses of shared code drive granularity integration. For example, infrastructurerelated cross-cutting functionality such as logging, auditing, authentication, authori‐
zation, and monitoring that all services use is not a good driver for putting services
back together or even moving back to a monolithic architecture. Some of the guide‐
lines for considering shared code as a granularity integrator are as follows:
Specific shared domain functionality
Shared domain functionality is shared code that contains business logic (as
opposed to infrastructure-related cross-cutting functionality). Our recommenda‐
tion is to consider this factor as a possible granularity integrator if the percentage
of shared domain code is relatively high. For example, suppose the common
(shared) code for a group of customer-related functionality (profile maintenance,
preference maintenance, and adding or removing comments) makes up over 40%
204
|
Chapter 7: Service Granularity
of the collective codebase. Breaking up the collective functionality into separate
services would mean that almost half of the source code is in a shared library
used only by those three services. In this example it might be wise to consider
keeping the collective customer-related functionality in a single consolidated ser‐
vice along with the shared code (particularly if the shared code changes fre‐
quently, as discussed next).
Frequent shared code changes
Regardless of the size of the shared library, frequent changes to shared function‐
ality require frequent coordinated changes to the services using that shared
domain functionality. While versioning can sometimes be used to help mitigate
coordinated changes, eventually services using that shared functionality will need
to adopt the latest version. If the shared code changes frequently, it might be wise
to consider consolidating the services using that shared code to help mitigate the
complex change coordination of multiple deployment units.
Defects that cannot be versioned
While versioning can help mitigate coordinated changes and allow for backward
compatibility and agility (the ability to respond quickly to change), at times cer‐
tain business functionality must be applied to all services at the same time (such
as a defect or a change in business rules). If this happens frequently, it might be
time to consider putting services back together to simplify the changes.
Data Relationships
Another trade-off in the balance between granularity disintegrators and integrators is
the relationship between the data that a single consolidated service uses as opposed to
the data that separate services would use. This integrator driver assumes that the data
resulting from breaking apart a service is not shared, but rather formed into tight
bounded contexts within each service to facilitate change control and support overall
availability and reliability.
Consider the example in Figure 7-16: a single consolidated service has three functions
(A, B, and C) and corresponding data table relationships. The solid lines pointing to
the tables represent writes to the tables (hence data ownership), and the dotted lines
pointing away from the tables represent read-only access to the table. Performing a
mapping operation between the functions and the tables reveals the results shown in
Table 7-1, where owner implies writes (and corresponding reads) and access implies
read-only access to a table not owned by that function.
Granularity Integrators
|
205
Figure 7-16. The database table relationships of a consolidated service
Table 7-1. Function-to-table mapping
Function Table 1 Table 2 Table 3 Table 4 Table 5 Table 6
A
B
C
owner owner
owner
owner
access
owner
access
owner
Assume that based on some of the disintegration drivers outlined in the prior section,
this service was broken into three separate services (one for each of the functions in
the consolidated service); see Figure 7-17. However, breaking apart the single consoli‐
dated service into three separate services now requires the corresponding data tables
to be associated with each service in a bounded context.
Notice at the top of Figure 7-17 that Service A owns tables 1, 2, 4, and 6 as part of its
bounded context; Service B owns table 3; and Service C owns table 5. However, notice
in the diagram that every operation in Service B requires access to data in table 5
(owned by Service C), and every operation in Service C requires access to data in
table 3 (owned by Service B). Because of the bounded context, Service B cannot sim‐
ply reach out and directly query table 5, nor can Service C directly query table 3.
To better understand the bounded context and why Service C cannot simply access
table 3, say Service B (which owns table 3) decides to make a change to its business
rules that requires a column to be removed from table 3. Doing so would break Ser‐
vice C and any other services using table 3. This is why the bounded context concept
is so important in highly distributed architectures like microservices. To resolve this
issue, Service B would have to ask Service C for its data, and Service C would have to
ask Service B for its data, resulting in back-and-forth interservice communication
between these services, as illustrated at the bottom of Figure 7-17.
206
|
Chapter 7: Service Granularity
Figure 7-17. Database table relationships impact service granularity
Based on the dependency of the data between services B and C, it would be wise to
consolidate those services into a single service to avoid the latency, fault tolerance,
and scalability issues associated with the interservice communication between these
services, demonstrating that relationships between tables can influence service granu‐
larity. We’ve saved this granularity integration driver for last because it is the one
granularity integration driver with the fewest number of trade-offs. While occasion‐
ally a migration from a monolithic system requires a refactoring of the way data is
organized, in most cases it isn’t feasible to reorganize database table entity relation‐
ships for the sake of breaking apart a service. We dive into the details about breaking
apart data in Chapter 6.
Granularity Integrators
|
207
Finding the Right Balance
Finding the right level of service granularity is hard. The secret to getting granularity
right is understanding both granularity disintegrators (when to break apart a service)
and granularity integrators (when to put them back together), and analyze the corre‐
sponding trade-offs between the two. As illustrated in the previous scenarios, this
requires an architect to not only identify the trade-offs, but also to collaborate closely
with business stakeholders to analyze those trade-offs and arrive at the appropriate
solution for service granularity.
Tables 7-2 and 7-3 summarize the drivers for disintegrators and integrators.
Table 7-2. Disintegrator drivers (breaking apart a service)
Disintegrator driver Reason for applying driver
Service scope
Code volatility
Scalability
Fault tolerance
Security access
Extensibility
Single-purpose services with tight cohesion
Agility (reduced testing scope and deployment risk)
Lower costs and faster responsiveness
Better overall uptime
Better security access control to certain functions
Agility (ease of adding new functionality)
Table 7-3. Integrator drivers (putting services back together)
Integrator driver
Reason for applying driver
Database transactions
Workflow
Shared code
Data relationships
Data integrity and consistency
Fault tolerance, performance, and reliability
Maintainability
Data integrity and correctness
Architects can use the drivers in these tables to form trade-off statements that can
then be discussed and resolved by collaborating with a product owner or business
sponsor.
Example 1:
Architect: “We want to break apart our service to isolate frequent code changes, but in
doing so we won’t be able to maintain a database transaction. Which is more important
based on our business needs—better overall agility (maintainability, testability, and
deployability), which translates to faster time-to-market, or stronger data integrity and
consistency?”
Project Sponsor: “Based on our business needs, I’d rather sacrifice a little bit slower
time-to-market to have better data integrity and consistency, so let’s leave it as a single
service for right now.”
208
|
Chapter 7: Service Granularity
Example 2:
Architect: “We need to keep the service together to support a database transaction
between two operations to ensure data consistency, but that means sensitive function‐
ality in the combined single service will be less secure. Which is more important based
on our business needs—better data consistency or better security?”
Project Sponsor: “Our CIO has been through some rough situations with regard to
security and protecting sensitive data, and it’s on the forefront of their mind and part
of almost every discussion. In this case, it’s more important to secure sensitive data, so
let’s keep the services separate and work out how we can mitigate some of the issues
with data consistency.”
Example 3:
Architect: “We need to break apart our payment service to provide better extensibility
for adding new payment methods, but that means we will have increased workflow
that will impact the responsiveness when multiple payment types are used for an order
(which happens frequently). Which is more important based on our business needs—
better extensibility within the payment processing, hence better agility and overall
time-to-market, or better responsiveness for making a payment?”
Project Sponsor: “Given that I see us adding only two, maybe three more payment
types over the next couple of years, I’d rather have us focus on the overall responsive‐
ness since the customer must wait for payment processing to be complete before the
order ID is issued.”
Sysops Squad Saga: Ticket Assignment Granularity
Monday, October 25 11:08
Once a trouble ticket has been created by a customer and accepted by the system, it
must be assigned to a Sysops Squad expert based on their skill set, location, and availability. Ticket assignment involves two main components—a Ticket Assignment component that determines which consultant should be assigned the job, and the Ticket
Routing component that locates the Sysops Squad expert, forwards the ticket to the
expert’s mobile device (via a custom Sysops Squad mobile app), and notifies the expert
via an SMS text message that a new ticket has been assigned.
The Sysops Squad development team was having trouble deciding whether these two components
(assignment and routing) should be implemented as a single consolidated service or two separate
services, as illustrated in Figure 7-18. The development team consulted with Addison (one of the
Sysops Squad architects) to help decide which option it should go with.
Sysops Squad Saga: Ticket Assignment Granularity
|
209
Figure 7-18. Options for ticket assignment and routing
“So you see,” said Taylen, “the ticket assignment algorithms are very complex, and therefore should
be isolated from the ticket routing functionality. That way, when those algorithms change, I don’t
have to worry about all of the routing functionality.”
“Yes, but how much change is there to those assignment algorithms?” asked Addison. “And how
much change do we anticipate in the future?”
“I apply changes to those algorithms at least two to three times a month. I read about volatilitybased decomposition, and this situation fits it perfectly,” said Taylen.
“But if we separated the assignment and routing functionality into two services, there would need to
be constant communication between them,” said Skyler. “Furthermore, assignment and routing are
really one function, not two.”
“No,” said Taylen, “they are two separate functions.”
“Hold on,” said Addison. “I see what Skyler means. Think about it a minute. Once an expert is found
that is available within a certain period of time, the ticket is immediately routed to that expert. If no
expert is available, the ticket goes back in the queue and waits until an expert can be found.”
“Yes, that’s right,” said Taylen.
210
|
Chapter 7: Service Granularity
“See,” said Skyler, “you cannot make a ticket assignment without routing it to the expert. So the two
functions are one.”
“No, no, no,” said Taylen. “You don’t understand. If an expert is seen to be available within a certain
amount of time, then that expert is assigned. Period. Routing is just a transport thing.”
“What happens in the current functionality if a ticket can’t be routed to the expert?” asked Addison.
“Then another expert is selected,” said Taylen.
“OK, so think about it a minute, Taylen,” said Addison. “If assignment and routing are two separate
services, then the routing service would have to then communicate back to the assignment service,
letting it know that the expert cannot be located and to pick another one. That’s a lot of coordination between the two services.”
“Yes, but they are still two separate functions, not one as Skyler is suggesting,” said Taylen.
“I have an idea,” said Addison. “Can we all agree that the assignment and routing are two separate
activities, but are tightly bound synchronously to each other? Meaning, one function cannot exist
without the other?”
“Yes,” both Taylen and Skyler replied.
“In that case,” said Addison, “let’s analyze the trade-offs. Which is more important—isolating the
assignment functionality for change control purposes, or combining assignment and routing into a
single service for better performance, error handling, and workflow control?”
“Well,” said Taylen, “when you put it that way, obviously the single service. But I still want to isolate
the assignment code.”
“OK,” said Addison, “in that case, how about we make three distinct architectural components in the
single service. We can delineate assignment, routing, and shared code with separate namespaces in
the code. Would that help?”
“Yeah,” said Taylen, “that would work. OK, you both win. Let’s go with a single service then.”
“Taylen,” said Addison, “it’s not about winning, it’s about analyzing the trade-offs to arrive at the most
appropriate solution; that’s all.”
With everyone agreeing to a single service for assignment and routing, Addison wrote the following
architecture decision record (ADR) for this decision:
ADR: Consolidated Service for Ticket Assignment and Routing
Context
Once a ticket is created and accepted by the system, it must be assigned to an expert and
then routed to that expert’s mobile device. This can be done through a single consolidated ticket assignment service or separate services for ticket assignment and ticket routing.
Sysops Squad Saga: Ticket Assignment Granularity
|
211
Decision
We will create a single consolidated ticket assignment service for the assignment and
routing functions of the ticket.
Tickets are immediately routed to the Sysops Squad expert once they are assigned, so
these two operations are tightly bound and dependent each other.
Both functions must scale the same, so there are no throughput differences between
these services, nor is back-pressure needed between these functions.
Since both functions are fully dependent on each other, fault tolerance is not a driver for
breaking these functions apart.
Making these functions separate services would require workflow between them, resulting in performance, fault tolerance, and possible reliability issues.
Consequences
Changes to the assignment algorithm (which occur on a regular basis) and changes to the
routing mechanism (infrequent change) would require testing and deployment of both
functions, resulting in increased testing scope and deployment risk.
Sysops Squad Saga: Customer Registration Granularity
Friday January 14, 13:15
Customers must register with the system to gain access to the Sysops Squad support
plan. During registration, customers must provide profile information (name, address,
business name if applicable, and so on), credit card information (which is billed on a
monthly basis), password and security question information, and a list of products purchased they would like to have covered under the Sysops Squad support plan.
Some members of the development team insisted that this should be a single consolidated Customer Service containing all of the customer information, yet other members
of the team disagreed and thought that there should be a separate service for each of these functions (a Profile service, Credit Card service, Password service, and a Supported Product service). Skyler, having prior experience in PCI and PII data, thought that the credit card and password
information should be a separate service from the rest, and hence only two services (a Profile service
containing profile and product information and a separate Customer Secure service containing
credit card and password information). These three options are illustrated in Figure 7-19.
212
|
Chapter 7: Service Granularity
Figure 7-19. Options for customer registration
Because Addison was busy with the core ticketing functionality, the development team asked for
Austen’s help in resolving this granularity issue. Anticipating this will not be an easy decision, particularly since it involved security, Austen scheduled a meeting with Parker, (the product owner), and
Sam, the Penultimate Electronics security expert to discuss these options.
“OK, so what can we do for you?” asked Parker.
“Well,” said Austen, “we are struggling with how many services to create for registering customers
and maintaining customer-related information, You see, there are four main pieces of data we are
dealing with here: profile info, credit card info, password info, and purchased product info.”
Sysops Squad Saga: Customer Registration Granularity
|
213
“Whoa, hold on now,” interrupted Sam. “You know that credit card and password information must
be secure, right?”
“Of course we know it has to be secure,” said Austen. “What we’re struggling with is the fact that
there’s a single customer registration API to the backend, so if we have separate services they all
have to be coordinated together when registering a customer, which would require a distributed
transaction.”
“What do you mean by that?” asked Parker.
“Well,” said Austen, “we wouldn’t be able to synchronize all of the data together as one atomic unit
of work.”
“That’s not an option,” said Parker. “All of the customer information is either saved in the database, or
it’s not. Let me put it another way. We absolutely cannot have the situation where we have a customer record without a corresponding credit card or password record. Ever.”
“OK, but what about securing the credit card and password information?” asked Sam. “Seems to me,
having separate services would allow much better security control access to that type of sensitive
information.”
“I think I may have an idea.” said Austen. “The credit card information is tokenized in the database,
right?”
“Tokenized and encrypted,” said Sam.
“Great. And the password information?” asked Austen.
“The same,” said Sam.
“OK,” said Austen, “so it seems to me that what we really need to focus on here is controlling access
to the password and credit card information separate from the other customer-related requests—
you know, like getting and updating profile information, and so on.”
“I think I see where you are coming from with your problem,” said Parker. “You’re telling me that if
you separate all of this functionality into separate services, you can better secure access to sensitive
data, but you cannot guarantee my all-or-nothing requirement. Am I right?”
“Exactly. That’s the trade-off,” said Austen.
“Hold on,” said Sam. “Are you using the Tortoise security libraries to secure the API calls?”
“Yes. We use those libraries not only at the API layer, but also within each service to control access
through the service mesh. So essentially it’s a double-check,” said Austen.
“Hmmm,” said Sam. “OK, I’m good with a single service providing you use the Tortoise security
framework.”
“Me too, providing we can still have the all-or-nothing customer registration process,” said Parker.
214
|
Chapter 7: Service Granularity
“Then I think we are all in agreement that the all-or-nothing customer registration is an absolute
requirement and we will maintain multilevel security access using Tortoise,” said Austen.
“Agreed,” said Parker.
“Agreed,” said Sam.
Parker noticed how Austen handled the meeting by facilitating the conversation rather than controlling it. This was an important lesson as an architect in identifying, understanding, and negotiating
trade-offs. Parker also better understood the difference between design versus architecture in that
security can be controlled through design (use of a custom library with special encryption) rather
than architecture (breaking up functionality into separate deployment units).
Based on the conversation with Parker and Sam, Austen made the decision that customer-related
functionality would be managed through a single consolidated domain service (rather than separately deployed services) and wrote the following ADR for this decision:
ADR: Consolidated Service for Customer-Related Functionality
Context
Customers must register with the system to gain access to the Sysops Squad support plan.
During registration, customers must provide profile information, credit card information,
password information, and products purchased. This can be done through a single consolidated customer service, a separate service for each of these functions, or a separate
service for sensitive and nonsensitive data.
Decision
We will create a single consolidated customer service for profile, credit card, password, and
products supported.
Customer registration and unsubscribe functionality requires a single atomic unit of work.
A single service would support ACID transactions to meet this requirement, whereas separate services would not.
Use of the Tortoise security libraries in the API layer and the service mesh will mitigate
security access risk to sensitive information.
Consequences
We will require the Tortoise security library to ensure security access in both the API gateway and the service mesh.
Because it’s a single service, changes to source code for profile info, credit card, password,
or products purchased will increase testing scope and increase deployment risk.
The combined functionality (profile, credit card, password, and products purchased) will
have to scale as one unit.
The trade-off discussed in a meeting with the product owner and security expert is transactionality versus security. Breaking the customer functionality into separate services
Sysops Squad Saga: Customer Registration Granularity
|
215
provides better security access, but doesn’t support the “all-or-nothing” database transaction required for customer registration or unsubscribing. However, the security concerns
are mitigated through the use the custom Tortoise security library.
216
| Chapter 7: Service Granularity
PART II
Putting Things Back Together
Attempting to divide a cohesive module would only result in increased coupling and
decreased readability.
—Larry Constantine
Once a system is broken apart, architects often find it necessary to stitch it back
together to make it work as one cohesive unit. As Larry Constantine so eloquently
infers in the preceding quote, it’s not quite as easy as it sounds, with lots of trade-offs
involved when breaking things apart.
In this second part of this book, we discuss various techniques for overcoming some
of the hard challenges associated with distributed architectures, including managing
service communication, contracts, distributed workflows, distributed transactions,
data ownership, data access, and analytical data.
Part I was about structure; Part II is about communication. Once an architect under‐
stands the structure and the decisions that lead to it, it’s time to think about how the
structural parts interact with each other.
CHAPTER 8
Reuse Patterns
Wednesday, February 2, 15:15
As the development team members worked on breaking apart the domain services,
they started running into disagreements about what to do with all the shared code and
shared functionality. Taylen, upset with what Skyler was doing with regard to the shared
code, walked over to Skyler’s desk.
“What in the world are you doing?” asked Taylen.
“I’m moving all of the shared code to a new workspace so we can create a shared DLL
from it,” replied Skyler.
“A single shared DLL?”
“That’s what I was planning,” said Skyler. “Most of the services will need this stuff anyway, so I’m
going to create a single DLL that all the services can use.”
“That’s the worst idea I’ve ever heard,” said Taylen. “Everyone knows you should have multiple shared
libraries in a distributed architecture!”
“Not in my opinion,” said Sydney. “Seems to me it’s much easier to manage a single shared library
DLL rather than dozens of them.”
“Given that I’m the tech lead for this application, I want you to split that functionality into separate
shared libraries.”
“OK, OK, I suppose I can move the all of the authorization into its own separate DLL if that would
make you happy,” said Skyler.
“What?” said Taylen. “The authorization code has to be a shared service, you know——not in a
shared library."”
219
“No,” said Skyler. “That code should be in a shared DLL.”
“What’s all the shouting about over there?” asked Addison.
“Taylen wants the authorization functionality to be in a shared service. That’s just crazy. I think it
should go in the common shared DLL,” said Skyler.
“No way,” said Taylen. “It’s got to be in its own separate shared service.”
“And,” said Skyler, “Taylen is insisting on having multiple shared libraries for the shared functionality
rather than a single shared library.”
“Tell you what,” said Addison. “Let’s go over the trade-offs of shared library granularity, and also go
over the trade-offs between a shared library and a shared service to see if we can resolve these
issues in a more reasonable and thoughtful manner.”
Code reuse is a normal part of software development. Common business domain
functionality, such as formatters, calculators, validators, and auditing, are typically
shared across multiple components, as is common infrastructure functionality, such
as security, logging, and metrics gathering. In most monolithic architectures, code
reuse is rarely given a second thought—it’s a matter of simply importing or autoinjecting shared class files. However, in distributed architectures, as shown in Figure
8-1, things get a bit more complicated, as questions arise about how to deal with
shared functionality.
Figure 8-1. Code reuse is a hard part of distributed architecture
Frequently within highly distributed architectures like microservices and serverless
environments, phrases like “reuse is abuse!” and “share nothing!” are touted by archi‐
tects in an attempt to reduce the amount of shared code within these types of archi‐
tectures. Architects in these environments have even been found to offer countering
advice to the famous DRY principle (Don’t repeat yourself) by using an opposing
acronym called WET (Write every time or Write everything twice).
220
| Chapter 8: Reuse Patterns
While developers should try to limit the amount of code reuse within distributed
architectures, it is nevertheless a fact of life in software development and must be
addressed, particularly in distributed architectures. In this chapter, we introduce sev‐
eral techniques for managing code reuse within a distributed architecture, including
replicating code, shared libraries, shared services, and sidecars within a service mesh.
For each of these options, we also discuss the pros, cons, and trade-offs of each
approach.
Code Replication
In code replication, shared code is copied into each service (or more specifically, each
service source code repository), as shown in Figure 8-2, thereby avoiding code shar‐
ing altogether. While it might sound crazy, this technique became popular in the early
days of microservices when a lot of confusion and misunderstanding arose about the
bounded context concept, hence the drive to create a “share nothing architecture.” In
theory, code replication seemed like a good approach at that time to reduce code
sharing, but in practice it quickly fell apart.
Figure 8-2. With replication, shared functionality is copied into each service
While code replication isn’t used much today, it nevertheless is still a valid technique
for addressing code reuse across multiple distributed services. This technique should
be approached with extreme caution for the obvious reason that if a bug is found in
the code or an important change to the code is needed, it would be very difficult and
time-consuming to update all services containing the replicated code.
At times, however, this technique can prove useful, particularly for highly static oneoff code that most (or all) services need. For example, consider the Java code in
Example 8-1 and the corresponding C# code in Example 8-2 that identifies the class
in the service that represents the service entry point (usually the restful API class
within a service).
Code Replication
|
221
Example 8-1. Source code defining a service entry point annotation (Java)
@Retention(RetentionPolicy.RUNTIME)
@Target(ElementType.TYPE)
public @interface ServiceEntrypoint {}
/* Usage:
@ServiceEntrypoint
public class PaymentServiceAPI {
...
}
*/
Example 8-2. Source code defining a service entry point attribute (C#)
[AttributeUsage(AttributeTargets.Class)]
class ServiceEntrypoint : Attribute {}
/* Usage:
[ServiceEntrypoint]
class PaymentServiceAPI {
...
}
*/
Note that the source code in Example 8-1 actually contains no functionality whatso‐
ever. The annotation is simply a marker (or tag) used to identify a particular class as
representing the service entry point. However, this simple annotation is very useful
for placing other metadata annotations about a particular service, including the ser‐
vice type, domain, bounded context, and so on; see Chapter 89 in 97 Things Every
Java Programmer Should Know by Kevlin Henney and Trisha Gee (O’Reilly) for a
description of these metadata custom annotations.
This kind of source code makes a good candidate for replication because it’s static and
doesn’t contain any bugs (and most likely will not in the future). If this were a unique
one-off class, it might be worth copying it into each service code repository rather
than creating a shared library for it. That said, we generally encourage investigating
the other code-sharing techniques presented in this chapter before opting for the
code replication technique.
While the replication technique preserves the bounded context, it does make it diffi‐
cult to apply changes if the code ever does need to be modified. Table 8-1 lists the
various trade-offs associated with this technique.
222
| Chapter 8: Reuse Patterns
Trade-Offs
Table 8-1. Trade-offs for the code replication technique
Advantages
Disadvantages
Preserves the bounded context
No code sharing
Difficult to apply code changes
Code inconsistency across services
No versioning capabilities across services
When to Use
The replication technique is a good approach when developers have simple static
code (like annotations, attributes, simple common utilities, and so on) that is either a
one-off class or code that is unlikely to ever change because of defects or functional
changes. However, as mentioned earlier, we encourage exploring other code-reuse
options before embracing the code replication technique.
When migrating from a monolithic architecture to a distributed one, we’ve also
found that the replication technique can sometimes work for common static utility
classes. For example, by replicating a Utility.cs C# class to all services, each service
can now remove (or enhance) the Utility.cs class to suit its particular needs, there‐
fore eliminating unnecessary code and allowing the utility class to evolve for each
specific context (similar to the tactical forking technique described in Chapter 3).
Again, the risk with this technique is that a defect or change is very difficult to propa‐
gate to all services because the code is duplicated for each service.
Shared Library
One of the most common techniques for sharing code is to use a shared library. A
shared library is an external artifact (such as a JAR file, DLL, and so on) containing
source code that is used by multiple services which is typically bound to the service at
compile time (see Figure 8-3). Although the shared library technique seems simple
and straightforward, it has its share of complexities and trade-offs, not the least of
which is shared library granularity and versioning.
Shared Library
|
223
Figure 8-3. With the shared library technique, common code is consolidated and shared
at compile time
Dependency Management and Change Control
Similar to service granularity (discussed in Chapter 7), there are trade-offs associated
with the granularity of a shared library. The two opposing forces that form trade-offs
with shared libraries are dependency management and change control.
Consider the coarse-grained shared library illustrated in Figure 8-4. Note that while
the dependency management is relatively straightforward (each service uses the sin‐
gle shared library), change control is not. If a change occurs to any of the class files in
the coarse-grained shared library, every service, whether it cares about the change or
not, must eventually adopt the change because of a version deprecation of the shared
library. This forces unnecessary retesting and redeployment of all the services using
that library, therefore significantly increasing the overall testing scope of a shared
library change.
Figure 8-4. Changes to coarse-grained shared libraries impact multiple services but keep
dependencies low
224
|
Chapter 8: Reuse Patterns
Breaking shared code into smaller functionality-based shared libraries (such as secu‐
rity, formatters, annotations, calculators, and so on) is better for change control and
overall maintainability, but unfortunately creates a mess in terms of dependency
management. As shown in Figure 8-5, a change in shared class C7 impacts only Ser‐
vice D and Service E, but managing the dependency matrix between shared libraries
and services quickly starts looking like a big ball of distributed mud (or what some
people refer to as a distributed monolith).
Figure 8-5. Changes to fine-grained shared libraries impact fewer services but increase
dependencies
The choice of shared library granularity may not matter much with only a few serv‐
ices, but as the number of services increases, so do the issues associated with change
control and dependency management. Just imagine a system with 200 services and 40
shared libraries—it would quickly become overly complex and unmaintainable.
Given these trade-offs of change control and dependency management, our advice is
to generally avoid large, coarse-grained shared libraries and strive for smaller, func‐
tionally partitioned libraries whenever possible, thus favoring change control over
dependency management. For example, carving off relatively static functionality such
as formatters and security (authentication and authorization) into their own shared
libraries isolates this static code, therefore reducing the testing scope and unnecessary
version deprecation deployments for other shared functionality.
Versioning Strategies
Our general advice about shared library versioning is to always use versioning! Ver‐
sioning your shared libraries provides not only backward compatibility, but also a
high level of agility—the ability to respond quickly to change.
To illustrate this point, consider a shared library containing common field validation
rules called Validation.jar that is used by 10 services. Suppose one of those services
Shared Library
|
225
needs an immediate change to one of the validation rules. By versioning the Valida‐
tion.jar file, the service needing the change can immediately incorporate the new Val‐
idation.jar version and be deployed to production right away, without any impact to
the other 9 services. Without versioning, all 10 services would have to be tested and
redeployed when making the shared library change, thereby increasing the amount of
time and coordination for the shared library change (hence less agility).
While the preceding advice may seem obvious, there are trade-offs and hidden com‐
plexity in versioning. As a matter of fact, versioning can be so complex that your
authors often think of versioning as the ninth fallacy of distributed computing: “ver‐
sioning is simple”
One of the first complexities of shared library versioning is communicating a version
change. In a highly distributed architecture with multiple teams, it is often difficult to
communicate a version change to a shared library. How do other teams know that
Validation.jar just increased to version 1.5? What were the changes? What services
are impacted? What teams are impacted? Even with the plethora of tools that manage
shared libraries, versions, and change documentation (such as JFrog Artifactory),
version changes must nevertheless be coordinated and communicated to the right
people at the right time.
Another complexity is the deprecation of older versions of a shared library—remov‐
ing those versions no longer supported after a certain date. Deprecation strategies
range from custom (for individual shared libraries) all the way to global (for all shared
libraries). And, not surprisingly, trade-offs are involved with both approaches.
Assigning a custom deprecation strategy to each shared library is usually the desired
approach because libraries change at different rates. For example, if a Security.jar
shared library doesn’t change often, maintaining only two or three versions is a rea‐
sonable strategy. However, if the Calculators.jar shared library changes weekly, main‐
taining only two or three versions means that all services using that shared library
will be incorporating a newer version on a monthly (or even weekly) basis—causing a
lot of unnecessary frequent retesting and redeployment. Therefore, maintaining 10
versions of Calculators.jar would be a much more reasonable strategy because of the
frequency of change. The trade-off of this approach, however, is that someone must
maintain and track the deprecation for each shared library. This can sometimes be a
daunting task and is definitely not for the faint of heart.
Because change is variable among the various shared libraries, the global deprecation
strategy, while simpler, is a less effective approach. The global deprecation strategy
dictates that all shared libraries, regardless of the rate of change, will not support
more than a certain number of backward versions (for example, four). While this is
easy to maintain and govern, it can cause significant churn--the constant retesting
and redeploying of services—just to maintain compatibility with the latest version of
226
| Chapter 8: Reuse Patterns
a frequently changed shared library. This can drive teams crazy and significantly
reduce overall team velocity and productivity.
Regardless of the deprecation strategy used, serious defects or breaking changes to
shared code invalidate any sort of deprecation strategy, causing all services to adopt
the latest version of a shared library at once (or within a very short period of time).
This is another reason we recommend keeping shared libraries as fine-grained as
appropriate and avoid the coarse-grained SharedStuff.jar type of libraries containing
all the shared functionality in the system.
One last word of advice regarding versioning: avoid the use of the LATEST version
when specifying which version of a library a service requires. It has been our experi‐
ence that services using the LATEST version experience issues when doing quick fixes
or emergency hot deployments into production, because something in the LATEST
version might be incompatible with the service, therefore causing additional develop‐
ment and testing effort for the team to release the service into production.
While the shared library technique allows changes to be versioned (therefore provid‐
ing a good level of agility for shared code changes), dependency management can be
difficult and messy. Table 8-2 lists various trade-offs associated with this technique.
Trade-Offs
Table 8-2. Trade-offs for the shared library technique
Advantages
Disadvantages
Ability to version changes
Shared code is compile-based, reducing runtime
errors
Good agility for code shared code changes
Dependencies can be difficult to manage
Code duplication in heterogeneous codebases
Version deprecation can be difficult
Version communication can be difficult
When To Use
The shared library technique is a good approach for homogeneous environments
where shared code change is low to moderate. The ability to version (although some‐
times complex) allows for good levels of agility when making shared code changes.
Because shared libraries are usually bound to the service at compile time, operational
characteristics such as performance, scalability, and fault tolerance are not impacted,
and the risk of breaking other services with a change to common code is low because
of versioning.
Shared Library
|
227
Shared Service
The primary alternative to using a shared library for common functionality is to use a
shared service instead. The shared service technique, illustrated in Figure 8-6, avoids
reuse by placing shared functionality in a separately deployed service.
Figure 8-6. With the shared service technique, common functionality is made available
at runtime through separate services
One distinguishing factor about the shared service technique is that the shared code
must be in the form of composition, not inheritance. While there is a lot of debate
about the use of composition over inheritance from a source code design standpoint
(see the Thoughtworks article “Composition vs. Inheritance: How to Choose” and
Martin Fowler’s article “Designed Inheritance”), architecturally composition versus
inheritance matters when choosing a code-reuse technique, particularly with the
shared services technique.
Back in the day, shared services were a common approach to address shared func‐
tionality within a distributed architecture. Changes to shared functionality no longer
require redeployment of services; rather, since changes are isolated to a separate ser‐
vice, they can be deployed without redeploying other services needing the shared
functionality. However, like everything in software architecture, many trade-offs are
associated with using shared services, including change risk, performance, scalability,
and fault tolerance.
228
|
Chapter 8: Reuse Patterns
Change Risk
Changing shared functionality using the shared service technique turns out to be a
double-edged sword. As illustrated in Figure 8-7, changing shared functionality is
simply a matter of modifying the shared code contained in a separate service (such as
a discount calculator), redeploying the service, and voila—the changes are now avail‐
able to all services, without having to retest and redeploy any other service needing
that shared functionality.
Figure 8-7. Shared functionality changes are isolated to only the shared service
If only life were that simple! The problem, of course, is that a change to a shared ser‐
vice is a runtime change, as opposed to a compile-based change with the shared library
technique. As a result, a “simple” change in a shared service can effectively bring
down an entire system, as illustrated in Figure 8-8.
Figure 8-8. Changes to a shared service can break other services at runtime
Shared Service
|
229
This necessarily brings to the forefront the topic of versioning. In the shared library
technique, versioning is managed through compile-time bindings, significantly
reducing risk associated with a change in a shared library. However, how does one
version a simple shared service change?
The immediate response, of course, is to use API endpoint versioning—in other
words, create a new endpoint containing each shared service change, as shown in
Example 8-3.
Example 8-3. Discount calendar with versioning for shared service endpoint
app/1.0/discountcalc?orderid=123
app/1.1/discountcalc?orderid=123
app/1.2/discountcalc?orderid=123
app/1.3/discountcalc?orderid=123
latest change -> app/1.4/discountcalc?orderid=123
Using this approach, each time a shared service changes, the team would create a new
API endpoint containing a new version of the URI. It’s not difficult to see the issues
that arise with this practice. First of all, services accessing the discount calculator ser‐
vice (or the corresponding configuration for each service) must change to point to
the correct version. Second, when should the team create a new API endpoint? What
about for a simple error message change? What about for a new calculation? Version‐
ing starts to become largely subjective at this point, and the services using the shared
service must still change to point to the correct endpoint.
Another problem with API endpoint versioning is that it assumes all access to the
shared service is through a RESTful API call going through a gateway or via point-topoint communication. However, in some cases, access to a shared service through
interservice communication is commonly done through other types of protocols such
as messaging and gRPC (in addition to a RESTful API call). This further complicates
the versioning strategy for a change, making it difficult to coordinate versions across
multiple protocols.
The bottom line is that with the shared service technique, changes to a shared service
are generally runtime in nature, and therefore carry much more risk than with shared
libraries. While versioning can help reduce this risk, it’s much more complex to apply
and manage than that of a shared library.
230
| Chapter 8: Reuse Patterns
Performance
Because services requiring the shared functionality must make an interservice call to
a shared service, performance is impacted because of network latency (and security
latency, assuming the endpoints to the shared service are secure). This trade-off,
shown in Figure 8-9, does not exist with the shared library technique when accessing
shared code.
Figure 8-9. Shared service introduces network and security latency
Use of gRPC can help mitigate some of the performance issues by significantly reduc‐
ing network latency, as can the use of asynchronous protocols like messaging. With
messaging, the service needing the shared functionality can issue a request through a
request queue, perform other work, and once needed, can retrieve the results through
a separate reply queue using a correlation ID (see Java Message Service, Second Edi‐
tion by Mark Richards et al. (O’Reilly) for more information about messaging
techniques).
Shared Service
|
231
Scalability
Another drawback of the shared service technique is that the shared service must
scale as services using the shared service scale. This can sometimes be a mess to man‐
age, particularly with multiple services concurrently accessing the same shared ser‐
vice. However, as illustrated in Figure 8-10, the shared library technique does not
have this issue because the shared functionality is contained within the service at
compile time.
Figure 8-10. Shared services must scale as dependent services scale
Fault Tolerance
While fault-tolerance issues can usually be mitigated through multiple instances of a
service, nevertheless it is a trade-off to consider when using the shared service techni‐
que. As illustrated in Figure 8-11, if the shared service becomes unavailable, services
requiring the shared functionality are rendered nonoperational until the shared ser‐
vice is available. The shared library technique does not have this issue since the
shared functionality is contained in the service at compile time, and therefore
accessed through standard method or function calls.
232
|
Chapter 8: Reuse Patterns
Figure 8-11. Shared services introduce fault-tolerance issues
While the shared service technique preserves the bounded context and is good for
shared code that changes frequently, operational characteristics such as performance,
scalability, and availability suffer. Table 8-3 lists the various trade-offs associated with
this technique.
Trade-Offs
Table 8-3. Trade-offs for the shared service technique
Advantages
Disadvantages
Good for high code volatility
No code duplication in
heterogeneous codebases
Preserves the bounded context
Versioning changes can be difficult
Performance is impacted due to latency
No static code sharing
Fault tolerance and availability issues due to service
dependency
Scalability and throughput issues due to service
dependency
Increased risk due to runtime changes
Shared Service
|
233
When to Use
The shared service technique is good to use in highly polyglot environments (those
with multiple heterogeneous languages and platforms), and also when shared func‐
tionality tends to change often. While changes in a shared service tend to be much
more agile overall than with the shared library technique, be careful of runtime sideeffects and risks to services needing the shared functionality.
Sidecars and Service Mesh
Perhaps the most common response to any question posed by an architect is “It
depends!” No issue in distributed architectures better illustrates this ambiguity better
than operational coupling.
One of the design goals of microservices architectures is a high degree of decoupling,
often manifested in the advice “Duplication is preferable to coupling.” For example,
let’s say that two Sysops Squad services need to pass customer information, yet the
domain-driven design bounded context insists that implementation details remain
private to the service. Thus, a common solution allows each service its own internal
representation of entities such as Customer, passing that information in loosely cou‐
pled ways such as name-value pairs in JSON. Notice that this allows each service to
change its internal representation at will, including the technology stack, without
breaking the integration. Architects generally frown on duplicating code because it
causes synchronization issues, semantic drift, and a host of other issues, but some‐
times forces exist that are worse than the problems of duplication, and coupling in
microservices often fits that bill. Thus, in microservices architecture, the answer to
the question of “should we duplicate or couple to some capability?” is likely duplicate,
whereas in another architecture style such as a service-based architecture, the correct
answer is likely couple. It depends!
When designing microservices, architects have resigned themselves to the reality of
implementation duplication to preserve decoupling. But what about the type of capa‐
bilities that benefit from high coupling? For example, consider common operational
capabilities such as monitoring, logging, authentication and authorization, circuit
breakers, and a host of other operational abilities that each service should have. But
allowing each team to manage these dependencies often descends into chaos. For
example, consider a company like Penultimate Electronics trying to standardize on a
common monitoring solution to make it easier to operationalize the various services.
Yet if each team is responsible for implementing monitoring for their service, how
can the operations team be sure they did? Also, what about issues such as unified
upgrades? If the monitoring tool needs to upgrade across the organization, how can
teams coordinate that?
234
| Chapter 8: Reuse Patterns
The common solution that has emerged in the microservices ecosystem over the last
few years solves this problem in an elegant way, by using the Sidecar pattern. This
pattern is based on a much earlier architecture pattern defined by Alistair Cockburn,
known as the hexagonal architecture, illustrated in Figure 8-12.
In this Hexagonal pattern, what we would now call the domain logic resides in the
center of the hexagon, which is surrounded by ports and adaptors to other parts of
the ecosystem (in fact, this pattern is alternately known as the Ports and Adaptors Pat‐
tern). While predating microservices by a number of years, this pattern has similari‐
ties to modern microservices, with one significant difference: data fidelity. The
hexagonal architecture treated the database as just another adaptor that can be plug‐
ged in, but one of the insights from DDD suggests that data schemas and transaction‐
ality should be inside the interior—like microservices.
Figure 8-12. The Hexagonal pattern separated domain logic from technical coupling
Sidecars and Service Mesh
|
235
The Sidecar pattern leverages the same concept as hexagonal architecture in that it
decouples the domain logic from the technical (infrastructure) logic. For example,
consider two microservices, as shown in Figure 8-13.
Figure 8-13. Two microservices that share the same operational capabilities
Here, each service includes a split between operational concerns (the larger compo‐
nents toward the bottom of the service) and domain concerns, pictured in the boxes
toward the top of the service labeled “domain.” If architects desire consistency in
operational capabilities, the separable parts go into a sidecar component, metaphori‐
cally named for the sidecar that attaches to motorcycles, whose implementation is
either a shared responsibility across teams or managed by a centralized infrastructure
group. If architects can assume that every service includes the sidecar, it forms a con‐
sistent operational interface across services, typically attached via a service plane,
shown in Figure 8-14.
236
| Chapter 8: Reuse Patterns
Figure 8-14. When each microservice includes a common component, architects can
establish links between them for consistent control
If architects and operations can safely assume that every service includes the sidecar
component (governed by fitness functions), it forms a service mesh, illustrated in Fig‐
ure 8-15. The boxes to the right of each service all interconnect, forming a “mesh.”
Having a mesh allows architects and DevOps to create dashboards, control opera‐
tional characteristics such as scale, and a host of other capabilities.
The Sidecar pattern allows governance groups like enterprise architects a reasonable
restraint over too many polyglot environments: one of the advantages of microservi‐
ces is a reliance on integration rather than a common platform, allowing teams to
choose the correct level of complexity and capabilities on a service-by-service basis.
However, as the number of platforms proliferates, unified governance becomes more
difficult. Therefore, teams often use the consistency of the service mesh as a driver to
support infrastructure and other cross-cutting concerns across multiple heterogene‐
ous platforms. For example, without a service mesh, if enterprise architects want to
unify around a common monitoring solution, then teams must build a sidecar per
platform that supports that solution.
Sidecars and Service Mesh
|
237
Figure 8-15. A service mesh is an operational link among services
The Sidecar pattern represents not only a way to decouple operational capabilities
from domains—it’s an orthogonal reuse pattern to address a specific kind of coupling
(see “Orthogonal Coupling” on page 238). Often, architectural solutions require sev‐
eral types of coupling, such as our current example of domain versus operational cou‐
pling. An orthogonal reuse pattern presents a way to reuse some aspect counter to
one or more seams in the architecture. For example, microservices architectures are
organized around domains, but operational coupling requires cutting across those
domains. A sidecar allows an architect to isolate those concerns in a cross-cutting,
but consistent, layer through the architecture.
Orthogonal Coupling
In mathematics, two lines are orthogonal if they intersect at right angles, which also
implies independence. In software architecture, two parts of an architecture may be
orthogonally coupled: two distinct purposes that must still intersect to form a complete
solution. The obvious example from this chapter is an operational concern such as
monitoring, which is necessary but independent from domain behavior, like catalog
checkout. Recognizing orthogonal coupling allows architects to find intersection
points that cause the least entanglement between concerns.
238
| Chapter 8: Reuse Patterns
While the Sidecar pattern offers a nice abstraction, it has trade-offs like all other
architectural approaches, shown in Table 8-4.
Trade-Offs
Table 8-4. Trade-offs for the Sidecar pattern / service mesh technique
Advantages
Disadvantages
Offers a consistent way to create isolated coupling
Allows consistent infrastructure coordination
Must implement a sidecar per platform
Sidecar component may grow large/
complex
Ownership per team, centralized, or some
combination
When to Use
The Sidecar pattern and service mesh offer a clean way to spread some sort of crosscutting concern across a distributed architecture, and can be used by more than just
operational coupling (see Chapter 14). It offers an architectural equivalent to the Dec‐
orator Design Pattern from the Gang of Four Design Patterns book (Addison Wesley)
—it allows an architect to “decorate” behavior across a distributed architecture inde‐
pendent of the normal connectivity.
Sysops Squad Saga: Common Infrastructure Logic
Thursday, February 10, 10:34
Sydney peeped into Taylen’s office on a foggy morning. “Hey, are you using the shared
Message Dispatch library?”
Taylen replied, “Yes, we’re trying to consolidate on that to get some consistency on message resolution.”
Sydney said, “OK, but now we’re getting double log messages—it looks like the library
writes to the logs, but our service also writes to the log. Is that as it should be?”
“No,” Taylen replied. “We definitely don’t want duplicate log entries. That just makes everything confusing. We should ask Addison about that.”
Consequently, Sydney and Taylen darkened Addison’s door. “Hey, do you have a minute?”
Addison replied, “Always for you—what’s up?”
Sysops Squad Saga: Common Infrastructure Logic
|
239
Sydney said, “We’ve been consolidating a bunch of our duplicated code into shared libraries, and
that’s working well—we’re getting better at identifying the parts that rarely change. But, now we’ve
hit the problem that brings us here—who is supposed to be writing log messages? Libraries, services, or something else? And, how can we make that consistent?”
Addison said, “We’ve bumped into operational shared behavior. Logging is just one of them. What
about monitoring, service discovery, circuit breakers, even some of our utility functions, like the JSON
toXML library that a few teams are sharing? We need a better way to handle this to prevent issues.
That’s why we’re in the process of implementing a service mesh with this common behavior in a
sidecar component.”
Sydney said, “I’ve read about sidecars and service mesh—it’s a way to share things across a bunch of
microservices, right?”
Addison said, “Sort of, but not all kinds of things. The intent of the service mesh and sidecar is to
consolidate operational coupling, not domain coupling. For example, just like in our case, we want
consistency for logging and monitoring across all our services, but don’t want each team to have to
worry about that. If we consolidate logging code into the common sidecar that every service implements, we can enforce consistency.”
Taylen asked, “Who owns the shared library? Shared responsibility across all the teams?”
Addison replied, “We thought about that, but we have enough teams now; we’ve built a shared
infrastructure team that is going to manage and maintain the sidecar component. They have built
the deployment pipeline to automatically test the sidecar once it’s been bound into the service with
a set of fitness functions.”
Sydney said, “So if we need to share libraries between services, just ask them to put it in the sidecar?”
Addison said, “Be careful—the sidecar isn’t meant to be used for just anything, only operational
coupling.”
“I’m not sure what that distinction is,” Taylen said.
“Operational coupling includes the things we’ve been discussing—logging, monitoring, service discovery, authentication and authorization, and so on. Basically, it covers all the plumbing parts of the
infrastructure that have no domain responsibility. But you should never put domain shared components, like the Address or Customer class, in the sidecar.”
Sydney asked, “But why? What if I need the same class definition in two services? Won’t putting it in
the sidecar make it available to both?”
Addison replied, “Yes, but now you are increasing coupling in exactly the way we try to avoid in
microservices. In most architectures, a single implementation of that service would be shared across
the teams that need it. However, in microservices, that creates a coupling point, tying several services together in an undesirable way—if one team changes the shared code, every team must coordinate with that change. However, the architects could decide to put the shared library in the
240
| Chapter 8: Reuse Patterns
sidecar—it is, after all a technical capability. Neither answer is unambiguously correct, making this an
architect decision and worthy of trade-off analysis. For example, if the Address class changes and
both services rely on it, they must both change—the definition of coupling. We handle those issues
with contracts. The other issue concerns size: we don’t want the sidecar to become the biggest part
of the architecture. For example, consider the JSONtoXML library we were discussing before. How
many teams use that?”
Taylen said, “Well, any team that has to integrate with the mainframe system for anything—probably
5 out of, what, 16 or 17 teams?”
Addison said, “Perfect. OK, what’s the trade-off of putting the JSONtoXML in the sidecar?”
Sydney answered, “Well, that means that every team automatically has the library and doesn’t have
to wire it in through dependencies.”
“And the bad side?” asked Addison.
“Well, adding it to the sidecar makes it bigger, but not by much—it’s a small library.” said Sydney.
“That’s the key trade-off for shared utility code—how many teams need it versus how much overhead does it add to every service, particularly ones that don’t need it.”
“And if less than one-half the teams use it, it’s probably not worth the overhead,” Sydney said.
“Right! So, for now, we’ll leave that out of the sidecar and perhaps reassess in the future,” said
Addison.
ADR: Using a Sidecar for Operational Coupling
Context
Each service in our microservices architecture requires common and consistent operational behavior; leaving that responsibility to each team introduces inconsistencies and
coordination issues.
Decision
We will use a sidecar component in conjunction with a service mesh to consolidate
shared operational coupling.
The shared infrastructure team will own and maintain the sidecar for service teams; service
teams act as their customers. The following services will be provided by the sidecar:
• Monitoring
• Logging
• Service discovery
• Authentication
• Authorization
Sysops Squad Saga: Common Infrastructure Logic
|
241
Consequences
Teams should not add domain classes to the sidecar, which encourages inappropriate
coupling.
Teams work with the shared infrastructure team to place shared, operational libraries in the
sidecar if enough teams require it.
Code Reuse: When Does It Add Value?
Many architects fail to properly assess trade-offs when they encounter some situa‐
tions, which isn’t necessarily a deficiency—many trade-offs become obvious only
after the fact.
Reuse is one of the most abused abstractions, because the general view in organiza‐
tions is that reuse represents a laudable goal that teams should strive for. However,
failing to evaluate all the trade-offs associated with reuse can lead to serious problems
within architecture.
The danger of too much reuse was one of the lessons many architects learned from
the early 20th century trend of orchestration-driven service-oriented architecture,
where one of the primary goals for many organizations was to maximize reuse.
Consider the scenario from an insurance company, illustrated in Figure 8-16.
Figure 8-16. Each domain within a large insurance company has a view of the customer
242
|
Chapter 8: Reuse Patterns
Each division in the company has some aspect of customers it cares about. Years ago,
architects were instructed to keep an eye out for this type of commonality; once dis‐
covered, the goal was to consolidate the organizational view of customer into a single
service, shown in Figure 8-17.
Figure 8-17. Unifying on a centralized Customer service
While the picture in Figure 8-17 may seem logical, it’s an architectural disaster for
two reasons. First, if all institutional information about a key entity like Customer
must reside in a single place, that entity must be complex enough to handle any
domain and scenario, making it difficult to use for simple things.
Secondly, though, it creates brittleness within the architecture. If every domain that
needs customer information must get it from a single place, when that place changes,
everything breaks. For example, in our example, what happens when CustomerSer
vice needs to add new capabilities on behalf of one of the domains? That change
could potentially impact every other domain, requiring coordination and testing to
ensure that the change hasn’t “rippled” throughout the architecture.
What architects failed to realize is that reuse has two important aspects; they got the
first one correct: abstraction. The way architects and developers discover candidates
for reuse is via abstraction. However, the second consideration is the one that deter‐
mines utility and value: rate of change.
Observing that some reuse causes brittleness begs the question about how that kind
of reuse differs from the kinds we clearly benefit from. Consider things that everyone
successfully reuses: operating systems, open source frameworks and libraries, and so
Code Reuse: When Does It Add Value?
|
243
on. What distinguishes those from assets that project teams build? The answer is slow
rate of change. We benefit from technical coupling, like operating systems and exter‐
nal frameworks, because they have a well-understood rate of change and update
cadence. Internal domain capabilities or quick-changing technical frameworks make
terrible coupling targets.
Reuse is derived via abstraction but operationalized by slow rate of
change.
Reuse via Platforms
Much press exists extolling the virtue of platforms within organizations, almost to the
point of semantic diffusion. However, most agree that the platform is the new target of
reuse within organizations, meaning that for each distinguishable domain capability,
the organization builds a platform with a well-defined API to hide the implementa‐
tion details.
Slow rate of change drives this reasoning. As we discuss in Chapter 13, an API can be
designed to be quite loosely coupled to callers, allowing for an aggressive internal rate
of change of implementation details without breaking the API. This, of course,
doesn’t protect the organization from changes to the semantics of the information it
must pass between domains, but by careful design of encapsulation and contracts,
architects can limit the amount of breaking change and brittleness in integration
architecture.
Sysops Squad Saga: Shared Domain Functionality
Tuesday, February 8, 12:50
With Addison’s approval, the development team had decided to split the core ticketing
functionality into three separate services: a customer-facing Ticket Creation service, a
Ticket Assignment service, and a Ticket Completion service. However, all three services
used common database logic (queries and updates) and shared a set of database tables
in the ticketing data domain.
Taylen wanted to create a shared data service that would contain the common database
logic, thus forming a database abstraction layer, as shown in Figure 8-18.
244
|
Chapter 8: Reuse Patterns
Figure 8-18. Option using a shared Ticket Data service for common database logic for
the Sysops Squad ticketing services
Skyler hated the idea and wanted to use a single shared library (DLL) that each service would
include as part of the build and deployment, as illustrated in Figure 8-19.
Figure 8-19. Option using a shared library for common database logic for the Sysops
Squad ticketing services
Both developers met with Addison to resolve this roadblock.
“So, Addison, what is your opinion? Should the shared database logic be in a shared data service or a
shared library?” asked Taylen.
“It’s not about opinions,” said Addison. “It’s about analyzing the trade-offs to arrive at the most
appropriate solution for the core shared ticketing database functionality. Let’s do a hypothesis-based
approach and hypothesize that the most appropriate solution is to use the shared data service.”
“Hold on,” said Skyler. “It’s simply not a good architectural solution for this problem.”
“Why?” asked Addison, prompting Skyler to start thinking in terms of trade-offs.
Sysops Squad Saga: Shared Domain Functionality
|
245
“First of all,” said Skyler, “all three services would need to make an interservice call to the shared data
service for every database query or update. We’re going take a serious performance hit if we do that.
Furthermore, if the shared data service goes down, all three of those services become
nonoperational.”
“So?” said Taylen. “It’s all backend functionality, so who cares? The backend functionality doesn’t
have to be that fast, and services come up fairly quickly if they fail.”
“Actually,” said Addison, “it’s not all backend functionality. Don’t forget, the Ticket Creation service is
customer facing, and it would be using the same shared data service as the backend ticketing
functionality.”
“Yeah, but most of the functionality is still backend,” said Taylen, with a little less confidence than
before.
“So far,” said Addison, “it looks like the trade-off for using the shared data service is performance and
fault tolerance for the ticketing services.”
“Let’s also not forget that any changes made to the shared data service are runtime changes. In
other words,” said Skyler, “if we make a change and deploy the shared data service, we could possibly break something.”
“That’s why we test,” said Taylen.
“Yeah, but if you want to reduce risk you would have to test all of the ticketing services for every
change to the shared data service, which increases testing time significantly. With a shared DLL, we
could version the shared library to provide backward compatibility,” said Skyler.
“OK, we will add increased risk for changes and increased testing effort to the trade-offs as well,” said
Addison. “Also, let’s not forget that we would have extra coordination from a scalability standpoint.
Every time we create more instances of the ticket creation service, we would have to make sure we
create more instances of the shared data service as well.”
“Let’s not keep focusing so much on the negatives.” said Taylen. “How about the positives of using a
shared data service?”
“OK,” said Addison, “let’s talk about the benefits of using a shared data service.”
“Data abstraction, of course,” said Taylen. “The services wouldn’t have to worry about any database
logic. All they would have to do is make a remote service call to the shared data service.”
“Any other benefits?” asked Addison.
“Well,” said Taylen, “I was going to say centralized connection pooling, but we would need multiple
instances anyway to support the customer ticket creation service. It would help, but it’s not a major
game changer since there are only three services without a lot of instances of each service. However,
change control would be so much easier with a shared data service. We wouldn’t have to redeploy
any of the ticketing services for database logic changes.”
246
|
Chapter 8: Reuse Patterns
“Let’s take a look at those shared class files in the repository and see historically how much change
there really is for that code,” said Addison.
Addison, Taylen, and Skyler all looked at the repository history for the shared data logic class files.
“Hmm…” said Taylen, “I thought there were a lot more changes to that code than what is showing
up in the repo. OK, so I guess the changes are fairly minimal for the shared database logic after all.”
Through the conversation of discussing trade-offs, Taylen started to realize that the negatives of a
shared service seemed to outweigh the positives, and there was no real compelling justification for
putting the shared database logic in a shared service. Taylen agreed to put the shared database logic
in a shared DLL, and Addison wrote an ADR for this architecture decision:
ADR: Use of a Shared Library for Common Ticketing Database Logic
Context
The ticketing functionality is broken into three services: Ticket Creation, Ticket Assignment,
and Ticket Completion. All three services use common code for the bulk of the database
queries and update statements. The two options are to use a shared library or create a
shared data service.
Decision
We will use a shared library for the common ticketing database logic.
Using a shared library will improve performance, scalability, and fault tolerance of the
customer-facing Ticket Creation service, as well as for the Ticket Assignment service.
We found that the common database logic code does not change much and is therefore
fairly stable code. Furthermore, change is less risky for the common database logic
because services would need to be tested and redeployed. If changes are needed, we will
apply versioning where appropriate so that not all services need to be redeployed when
the common database logic changes.
Using a shared library reduces service coupling and eliminates additional service dependencies, HTTP traffic, and overall bandwidth.
Consequences
Changes to the common database logic in the shared DLL will require the ticketing services to be tested and deployed, therefore reducing overall agility for common database
logic for the ticketing functionality.
Service instances will need to manage their own database connection pool.
Sysops Squad Saga: Shared Domain Functionality
|
247
CHAPTER 9
Data Ownership and
Distributed Transactions
Friday, December 10 09:12
While the database team worked on decomposing the monolithic Sysops Squad database, the Sysops Squad development team, along with Addison, the Sysops Squad architect, started to work on forming bounded contexts between the services and the data,
assigning table ownership to services in the process.
“Why did you add the expert profile table to the bounded context of the Ticket Assignment service?” asked Addison.
“Because,” said Sydney, “the ticket assignment relies on that table for the assignment algorithms. It
constantly queries that table to get the expert’s location and skills information.”
“But it only does queries to the expert table,” said Addison. “The User Maintenance service contains
the functionality to perform database updates to maintain that information. Therefore, it seems to
me the expert profile table should be owned by the User Maintenance service and put within that
bounded context.”
“I disagree,” said Sydney. “We simply cannot afford for the assignment service to make remote calls
to the User Maintenance service for every query it needs. It simply won’t work.”
“In that case, how to you see updates occurring to the table when an expert acquires a new skill or
changes their service location? And what about when we hire a new expert?” asked Addison. “How
would that work?”
“Simple,” said Sydney. “The User Maintenance service can still access the expert table. All it would
need to do is connect to a different database. What’s the big deal about that?”
249
“Don’t you remember what Dana said earlier? It’s OK for multiple services to connect to the same
database schema, but it’s not OK for a service to connect to multiple databases or schemas. Dana
said that was a no-go and would not allow that to happen,” said Addison.
“Oh, right, I forgot about that rule. So what do we do?” asked Sydney. “We have one service that
needs to do occasional updates, and an entirely different service in an entirely different domain to
do frequent reads from the table.”
“I don’t know what the right answer is,” said Addison. “Clearly this is going to require more collaboration between the database team and us to figure these things out. Let me see if Dana can provide
any advice on this.”
Once data is pulled apart, it must be stitched back together to make the system work.
This means figuring out which services own what data, how to manage distributed
transactions, and how services can access data they need (but no longer own). In this
chapter, we explore the ownership and transactional aspects of putting distributed
data back together.
Assigning Data Ownership
After breaking apart data within a distributed architecture, an architect must deter‐
mine which services own what data. Unfortunately, assigning data ownership to a ser‐
vice is not as easy as it sounds, and becomes yet another hard part of software
architecture.
The general rule of thumb for assigning table ownership states that services that per‐
form write operations to a table own that table. While this general rule of thumb
works well for single ownership (only one service ever writes to a table), it gets messy
when teams have joint ownership (multiple services do writes to the same table) or
even worse, common ownership (most or all services write to the table).
The general rule of thumb for data ownership is that the service
that performs write operations to a table is the owner of that table.
However, joint ownership makes this simple rule complex!
To illustrate some of the complexities with data ownership, consider the example
illustrated in Figure 9-1 showing three services: a Wishlist Service that manages all of
the customer wish lists, a Catalog Service that maintains the product catalog, and an
Inventory Service that maintains the inventory and restocking functionality for all
products in the product catalog.
250
| Chapter 9: Data Ownership and Distributed Transactions
Figure 9-1. Once data is broken apart, tables must be assigned to services that own them
To further complicate matters, notice that the Wishlist Service writes to both the
Audit table and the Wishlist table, the Catalog Service writes to the Audit table and
the Product table, and the Inventory Service writes to the Audit table and the Product
table. Suddenly, this simple real-world example makes assigning data ownership a
complex and confusing task.
In this chapter, we unravel this complexity by discussing the three scenarios encoun‐
tered when assigning data ownership to services (single ownership, common owner‐
ship, and joint ownership), and exploring techniques for resolving these scenarios,
using Figure 9-1 as a common reference point.
Single Ownership Scenario
Single table ownership occurs when only one service writes to a table. This is the most
straightforward of the data ownership scenarios and is relatively easy to resolve.
Referring back to Figure 9-1, notice that the Wishlist table has only a single service
that writes to it—the Wishlist Service.
In this scenario, it is clear that the Wishlist Service should be the owner of the Wish‐
list table (regardless of other services that need read-only access to the Wishlist table),
see Figure 9-2. Notice that on the right side of this diagram, the Wishlist table
becomes part of the bounded context of the Wishlist Service. This diagramming tech‐
nique is an effective way to indicate table ownership and the bounded context formed
between the service and its corresponding data.
Single Ownership Scenario
|
251
Figure 9-2. With single ownership, the service that writes to the table becomes the table
owner
Because of the simplicity of this scenario, we recommend addressing single table
ownership relationships first to clear the playing field in order to better address the
more complicated scenarios that arise: common ownership and joint ownership.
Common Ownership Scenario
Common table ownership occurs when most (or all) of the services need to write to
the same table. For example, Figure 9-1 shows that all services (Wishlist, Catalog, and
Inventory) need to write to the Audit table to record the action performed by the
user. Since all services need to write to the table, it’s difficult to determine who should
actually own the Audit table. While this simple example includes only three services,
imagine a more realistic example where potentially hundreds (or even thousands) of
services must write to the same Audit table.
The solution of simply putting the Audit table in a shared database or shared schema
that is used by all services unfortunately reintroduces all of the data-sharing issues
described at the beginning of Chapter 6, including change control, connection starva‐
tion, scalability, and fault tolerance. Therefore, another solution is needed to solve
common data ownership.
A popular technique for addressing common table ownership is to assign a dedicated
single service as the primary (and only) owner of that data, meaning only one service
is responsible for writing data to the table. Other services needing to perform write
actions would send information to the dedicated service, which would then perform
the actual write operation on the table.
252
|
Chapter 9: Data Ownership and Distributed Transactions
If no information or acknowledgment is needed by services sending the data, services
can use persisted queues for asynchronous fire-and-forget messaging. Alternatively, if
information needs to be returned to the caller based on a write action (such as return‐
ing a confirmation number or database key), services can use something like REST,
gRPC, or request-reply messaging (pseudosynchronous) for a synchronous call.
Coming back to the Audit table example, notice in Figure 9-3 that the architect cre‐
ated a new Audit Service and assigned it ownership of the Audit table, meaning it is
the only service that performs read or write actions on the table. In this example,
since no return information is needed, the architect used asynchronous fire-andforget messaging with a persistent queue so that the Wishlist Service, Catalog Service,
and Inventory Service don’t need to wait for the audit record to be written to the
table. Making the queue persistent (meaning the message is stored on disk by the
broker) provides guaranteed delivery in the event of a service or broker failure and
helps ensure that no messages are lost.
Figure 9-3. Common ownership uses a dedicated service owner
In some cases, it may be necessary for services to read common data they don’t own.
These read-only access techniques are described in detail in Chapter 10.
Joint Ownership Scenario
One of the more common (and complex) scenarios involving data ownership is joint
ownership, which occurs when multiple services perform write actions on the same
table. This scenario differs from the prior common ownership scenario in that with
joint ownership, only a couple of services within the same domain write to the same
table, whereas with common ownership, most or all of the services perform write
Joint Ownership Scenario
|
253
operations on the same table. For example, notice in Figure 9-1 that all services per‐
form write operations on the Audit table (common ownership), whereas only the
Catalog and Inventory services perform write operations on the Product table (joint
ownership).
Figure 9-4 shows the isolated joint ownership example from Figure 9-1. The Catalog
Service inserts new products into the table, removes products no longer offered, and
updates static product information as it changes, whereas the Inventory Service is
responsible for reading and updating the current inventory for each product as prod‐
ucts are queried, sold, or returned.
Figure 9-4. Joint ownership occurs when multiple services within the same domain per‐
form write operations on the same table
Fortunately, several techniques exist to address this type of ownership scenario—the
table split technique, the data domain technique, the delegation technique, and the
service consolidation technique. Each is discussed in detail in the following sections.
Table Split Technique
The table split technique breaks a single table into multiple tables so that each service
owns a part of the data it’s responsible for. This technique is described in detail in the
book Refactoring Databases and in the companion website.
To illustrate the table split technique, consider the Product table example illustrated
in Figure 9-4. In this case, the architect or developer would first create a separate
Inventory table containing the product ID (key) and the inventory count (number of
items available), pre-populate the Inventory table with data from the existing Product
table, then finally remove the inventory count column from the Product table. The
source listing in Example 9-1 shows how this technique might be implemented using
data definition language (DDL) in a typical relational database.
254
| Chapter 9: Data Ownership and Distributed Transactions
Example 9-1. DDL source code for splitting up the Product table and moving inventory
counts to a new Inventory table
CREATE TABLE Inventory
(
product_id VARCHAR(10),
inv_cnt INT
);
INSERT INTO Inventory VALUES (product_id, inv_cnt)
AS SELECT product_id, inv_cnt FROM Product;
COMMIT;
ALTER TABLE Product DROP COLUMN inv_cnt;
Splitting the database table moves the joint ownership to a single table ownership sce‐
nario: the Catalog Service owns the data in the Product table, and the Inventory Ser‐
vice owns the data in the Inventory table. However, as shown in Figure 9-5, this
technique requires communication between the Catalog Service and Inventory Ser‐
vice when products are created or removed to ensure the data remains consistent
between the two tables.
Figure 9-5. Joint ownership can be addressed by breaking apart the shared table
For example, if a new product is added, the Catalog Service generates a product ID
and inserts the new product into the Product table. The Catalog Service then must
send that new product ID (and potentially the initial inventory counts) to the Inven‐
tory Service. If a product is removed, the Catalog Service first removes the product
from the Product table, then must notify the Inventory Service to remove the inven‐
tory row from the Inventory table.
Synchronizing data between split tables is not a trivial matter. Should communication
between the Catalog Service and the Inventory Service be synchronous or asynchro‐
nous? What should the Catalog Service do when adding or removing a product and
finding that the Inventory Service is not available? These are hard questions to
answer, and are usually driven by the traditional availability verses consistency
Joint Ownership Scenario
|
255
trade-off commonly found in distributed architectures. Choosing availability means
that it’s more important that the Catalog Service always be able to add or remove
products, even though a corresponding inventory record may not be created in the
Inventory table. Choosing consistency means that it’s more important that the two
tables always remain in sync with each other, which would cause a product creation
or removal operation to fail if the Inventory Service is not available. Because network
partitioning is necessary in distributed architectures, the CAP theorem states that
only one of these choices (consistency or availability) is possible.
The type of communication protocol (synchronous versus asynchronous) also mat‐
ters when splitting a table. Does the Catalog Service require a confirmation that the
corresponding Inventory record is added when creating a new product? If so, then
synchronous communication is required, providing better data consistency at the
sacrifice of performance. If no confirmation is required, the Catalog Service can use
asynchronous fire-and-forget communication, providing better performance at the
sacrifice of data consistency. So many trade-offs to consider!
Table 9-1 summarizes the trade-offs associated with the table split technique for joint
ownership.
Trade-Offs
Table 9-1. Joint ownership table split technique trade-offs
Advantages
Disadvantages
Preserves bounded context
Single data ownership
Tables must be altered and restructured
Possible data consistency issues
No ACID transaction between table updates
Data synchronization is difficult
Data replication between tables may occur
Data Domain Technique
Another technique for joint ownership is to create a shared data domain. This is
formed when data ownership is shared between the services, thus creating multiple
owners for the table. With this technique, the tables shared by the same services are
put into the same schema or database, therefore forming a broader bounded context
between the services and the data.
256
|
Chapter 9: Data Ownership and Distributed Transactions
Notice that Figure 9-6 looks close to the original diagram in Figure 9-4 with one
noticeable difference—the data domain diagram has the Product table in a separate
box outside the context of each owning service. This diagramming technique makes
it clear that the table is not owned by or part of the bounded context of either service,
but rather shared between them in a broader bounded context.
Figure 9-6. With joint ownership, services can share data by using the data domain tech‐
nique (shared schema)
While data sharing is generally discouraged in distributed architectures (particularly
with microservices), it does resolve some of the performance, availability, and data
consistency issues found in other joint ownership techniques. Because the services
are not dependent on each other, the Catalog Service can create or remove products
without needing to coordinate with the Inventory Service, and the Inventory Service
can adjust inventory without needing the Catalog Service. Both services become
completely independent from each other.
When choosing the data domain technique, always reevaluate why
separate services are needed since the data is common to each of
the services. Justifications might include scalability differences,
fault-tolerance needs, throughput differences, or isolating code vol‐
atility (see Chapter 7).
Unfortunately, sharing data in a distributed architecture introduces a number of
issues, the first of these being increased effort for changes made to the structure of the
data (such as changing the schema of a table). Because a broader bounded context is
formed between the services and the data, changes to the shared table structures may
require those changes to be coordinated among multiple services. This increases
development effort, testing scope, and deployment risk.
Joint Ownership Scenario
|
257
Another issue with the data domain technique with regard to data ownership is con‐
trolling which services have write responsibility to what data. In some cases, this
might not matter, but if it’s important to control write operations to certain data,
additional effort is required to apply specific governance rules to maintain specific
table or column write ownership.
Table 9-2 summarizes the trade-offs associated with the data domain technique for
the joint ownership scenario.
Trade-Offs
Table 9-2. Joint ownership data-domain technique trade-offs
Advantages
Disadvantages
Good data access performance
No scalability and throughput issues
Data remains consistent
No service dependency
Data schema changes involve more services
Increased testing scope for data schema changes
Data ownership governance (write responsibility)
Increased deployment risk for data schema changes
Delegate Technique
An alternative method for addressing the joint ownership scenario is the delegate
technique. With this technique, one service is assigned single ownership of the table
and becomes the delegate, and the other service (or services) communicates with the
delegate to perform updates on its behalf.
One of the challenges of the delegate technique is knowing which service to assign as
the delegate (the sole owner of the table). The first option, called primary domain pri‐
ority, assigns table ownership to the service that most closely represents the primary
domain of the data—in other words, the service that does most of the primary entity
CRUD operations for the particular entity within that domain. The second option,
called operational characteristics priority, assigns table ownership to the service need‐
ing higher operational architecture characteristics, such as performance, scalability,
availability, and throughput.
To illustrate these two options and the corresponding trade-offs associated with each,
consider the Catalog Service and Inventory Service joint ownership scenario shown
in Figure 9-4. In this example, the Catalog Service is responsible for creating, updat‐
ing, and removing products, as well as retrieving product information; the Inventory
Service is responsible for retrieving and updating product inventory count as well as
for knowing when to restock if inventory gets too low.
258
| Chapter 9: Data Ownership and Distributed Transactions
With the primary domain priority option, the service that performs most of the
CRUD operations on the main entity becomes the owner of the table. As illustrated in
Figure 9-7, since the Catalog Service performs most of the CRUD operations on
product information, the Catalog Service would be assigned as the single owner of
the table. This means that the Inventory service must communicate with the Catalog
Service to retrieve or update inventory counts since it doesn’t own the table.
Figure 9-7. Table ownership is assigned to the Catalog service because of domain priority
Like the common ownership scenario described earlier, the delegate technique always
forces interservice communication between the other services needing to update the
data. Notice in Figure 9-7 that the Inventory Service must send inventory updates
through some sort of remote access protocol to the Catalog Service so that it can per‐
form the inventory updates and reads on behalf of the Inventory Service. This com‐
munication can either be synchronous or asynchronous. As always in software
architecture, more trade-off analysis to consider.
With synchronous communication, the Inventory Service must wait for the inventory
to be updated by the Catalog Service, which impacts overall performance but ensures
data consistency. Using asynchronous communication to send inventory updates
makes the Inventory Service perform much faster, but the data is only eventually con‐
sistent. Furthermore, with asynchronous communication, because an error can occur
in the Catalog Service while trying to update inventory, the Inventory Service has no
guarantee that the inventory was ever updated, impacting data integrity as well.
With the operational characteristics priority option, the ownership roles would be
reversed because inventory updates occur at a much faster rate than static product
data. In this case, table ownership would be assigned to the Inventory Service, the jus‐
tification being that updating product inventory is a part of the frequent real-time
transactional processing of purchasing products as opposed to the more infrequent
administrative task of updating product information or adding and removing prod‐
ucts (see Figure 9-8).
Joint Ownership Scenario
|
259
With this option, frequent updates to inventory counts can use direct database calls
rather than remote access protocols, therefore making inventory operations much
faster and more reliable. In addition, the most volatile data (inventory count) is kept
highly consistent.
Figure 9-8. Table ownership is assigned to the Inventory Service because of operational
characteristics priority
However, one major problem with the diagram illustrated in Figure 9-8 is that of
domain management responsibility. The Inventory Service is responsible for manag‐
ing product inventory, not the database activity (and corresponding error handling)
for adding, removing, and updating static product information. For this reason, we
usually recommend the domain priority option, and leveraging things like a replica‐
ted in-memory cache or a distributed cache to help address performance and faulttolerance issues.
Regardless of which service is assigned as the delegate (sole table owner), the delegate
technique has some disadvantages, the biggest being service coupling and the need
for interservice communication. This in turn leads to other issues for nondelegate
services, including the lack of an atomic transaction when performing write opera‐
tions, low performance due to network and processing latency, and low fault toler‐
ance. Because of these issues, the delegate technique is generally better suited for
database write scenarios that do not require atomic transactions and that can tolerate
eventual consistency through asynchronous communications.
260
|
Chapter 9: Data Ownership and Distributed Transactions
Table 9-3 summarizes the overall trade-offs of the delegate technique.
Trade-Offs
Table 9-3. Joint ownership delegate technique trade-offs
Advantages
Disadvantages
Forms single table ownership
Good data schema change control
Abstracts data structures from other services
High level of service coupling
Low performance for nonowner writes
No atomic transaction for nonowner writes
Low fault tolerance for nonowner services
Service Consolidation Technique
The delegate approach discussed in the prior section highlights the primary issue
associated with joint ownership—service dependency. The service consolidation tech‐
nique resolves service dependency and addresses joint ownership by combining mul‐
tiple table owners (services) into a single consolidated service, thus moving joint
ownership into a single ownership scenario (see Figure 9-9).
Figure 9-9. Table ownership is resolved by combining services
Like the data domain technique, this technique resolves issues associated with service
dependencies and performance, while at the same time addressing the joint owner‐
ship problem. However, like the other techniques, it has its share of trade-offs as well.
Combining services creates a more coarse-grained service, thereby increasing the
overall testing scope as well as overall deployment risk (the chance of breaking some‐
thing else in the service when a new feature is added or a bug is fixed). Consolidating
services might also impact overall fault tolerance since all parts of the service fail
together.
Service Consolidation Technique
|
261
Overall scalability is also impacted when using the service consolidation technique
because all parts of the service must scale equally, even though some functionality
might not need to scale at the same level as other functionality. For example, in Figure
9-9, the catalog maintenance functionality (what used to be in a separate Catalog ser‐
vice) must unnecessarily scale to meet the high demands of the inventory retrieval
and update functionality.
Table 9-4 summarizes the overall trade-offs of the service consolidation technique.
Trade-Offs
Table 9-4. Joint ownership service consolidation technique trade-offs
Advantages
Disadvantages
Preserves atomic transactions
Good overall performance
More coarse-grained scalability
Less fault tolerance
Increased deployment risk
Increased testing scope
Data Ownership Summary
Figure 9-10 shows the resulting table ownership assignments from Figure 9-1 after
applying the techniques described in this section. For the single table scenario involv‐
ing the Wishlist Service, we simply assigned ownership to the Wishlist Service, form‐
ing a tight bounded context between the service and the table. For the common
ownership scenario involving the audit table, we created a new Audit Service, with all
other services sending an asynchronous message to a persisted queue. Finally, for the
more complex joint ownership scenario involving the product table with the Catalog
Service and Inventory Service, we chose to use the delegate technique, assigning sin‐
gle ownership of the product table to the Catalog Service, with the Inventory Service
sending update requests to the Catalog Service.
Once table ownership has been assigned to services, an architect must then validate
the table ownership assignments by analyzing business workflows and their corre‐
sponding transaction requirements.
262
|
Chapter 9: Data Ownership and Distributed Transactions
Figure 9-10. Resulting data ownership using delegate technique for joint ownership
Distributed Transactions
When architects and developers think about transactions, they usually think about a
single atomic unit of work where multiple database updates are either committed
together or all rolled back when an error occurs. This type of atomic transaction is
commonly referred to as an ACID transaction. As noted in Chapter 6, ACID is an
acronym describing the basic properties of an atomic single-unit-of-work database
transaction: atomicity, consistency, isolation, and durability.
To understand how distributed transactions work and the trade-offs involved with
using a distributed transaction, it’s necessary to fully understand the four properties
of an ACID transaction. We firmly believe that without an understanding of ACID
transactions, an architect cannot perform the necessary trade-off analysis for know‐
ing when (and when not to) use a distributed transaction. Therefore, we will dive into
the details of an ACID transaction first, then describe how they differ from dis‐
tributed transactions.
Distributed Transactions
|
263
Atomicity means a transaction must either commit or roll back all of its updates in a
single unit of work, regardless of the number of updates made during that transac‐
tion. In other words, all updates are treated as a collective whole, so all changes either
get committed or get rolled back as one unit. For example, assume registering a cus‐
tomer involves inserting customer profile information into a Customer Profile table,
inserting credit card information into a Wallet table, and inserting security-related
information into a Security table. Suppose the profile and credit card information are
successfully inserted, but the security information insert fails. With atomicity, the
profile and credit card inserts would be rolled back, keeping the database tables in
sync.
Consistency means that during the course of a transaction, the database would never
be left in an inconsistent state or violate any of the integrity constraints specified in
the database. For example, during an ACID transaction, the system cannot add a
detail record (such as an item) without first adding the corresponding summary
record (such as an order). Although some databases defer this check until commit
time, in general programmers cannot violate consistency constraints such as a
foreign-key constraint during the course of a transaction.
Isolation refers to the degree to which individual transactions interact with each
other. Isolation protects uncommitted transaction data from being visible to other
transactions during the course of the business request. For example, during the
course of an ACID transaction, when the customer profile information is inserted
into the Customer Profile table, no other services outside of the ACID transaction
scope can access the newly inserted information until the entire transaction is
committed.
Durability means that once a successful response from a transaction commit occurs,
it is guaranteed that all data updates are permanent, regardless of further system
failures.
To illustrate an ACID transaction, suppose a customer registering for the Sysops
Squad application enters all of their profile information, the electronic products they
want covered under the support plan, and their billing information on a single user
interface screen. This information is then sent to the single Customer Service, as
shown in Figure 9-11, which then performs all of the database activity associated with
the customer registration business request.
264
|
Chapter 9: Data Ownership and Distributed Transactions
Figure 9-11. With ACID transactions, an error on the billing insert causes a rollback to
the other table inserts
First, notice that with an ACID transaction, because an error occurred when trying to
insert the billing information, both the profile information and support contract
information that were previously inserted are now rolled back (that’s the atomicity
and consistency parts of ACID). While not illustrated in the diagram, data inserted
into each table during the course of the transaction is not visible to other requests
(that’s the isolation part of ACID).
Note that ACID transactions can exist within the context of each service in a dis‐
tributed architecture, but only if the corresponding database supports ACID proper‐
ties as well. Each service can perform its own commits and rollbacks to the tables it
owns within the scope of the atomic business transaction. However, if the business
request spans multiple services, the entire business request itself cannot be an ACID
transaction—rather, it becomes a distributed transaction.
Distributed transactions occur when an atomic business request containing multiple
database updates is performed by separately deployed remote services. Notice in Fig‐
ure 9-12 that the same request for a new customer registration (denoted by the laptop
image representing the customer making the request) is now spread across three sep‐
arately deployed services—a Customer Profile Service, a Support Contract Service,
and a Billing Payment Service.
Distributed Transactions
|
265
Figure 9-12. Distributed transactions do not support ACID properties
As you can see, distributed transactions do not support ACID properties.
Atomicity is not supported because each separately deployed service commits its own
data and performs only one part of the overall atomic business request. In a dis‐
tributed transaction, atomicity is bound to the service, not the business request (such
as customer registration).
Consistency is not supported because a failure in one service causes the data to be out
of sync between the tables responsible for the business request. As shown in Figure
9-12, since the Billing Payment Service insert failed, the Profile table and Contract
table are now out of sync with the Billing table (we’ll show how to address these
issues later in this section). Consistency is also impacted because traditional relational
database constraints (such as a foreign key always matching a primary key) cannot be
applied during each individual service commit.
Isolation is not supported because once the Customer Profile Service inserts the pro‐
file data in the course of a distributed transaction to register a customer, that profile
information is available to any other service or request, even though the customer
registration process (the current transaction) hasn’t completed.
Durability is not supported across the business request—it is supported for only each
individual service. In other words, any individual commit of data does not ensure
that all data within the scope of the entire business transaction is permanent.
266
|
Chapter 9: Data Ownership and Distributed Transactions
Instead of ACID, distributed transactions support something called BASE. In chemis‐
try, an acid substance and a base substance are exactly the opposite. The same is true
with atomic and distributed transactions—ACID transactions are opposite of BASE
transactions. BASE describes the properties of a distributed transaction: basic availa‐
bility, soft state, and eventual consistency.
Basic availability (the “BA” part of BASE) means that all of the services or systems in
the distributed transaction are expected to be available to participate in the dis‐
tributed transaction. While asynchronous communication can help decouple services
and address availability issues associated with the distributed transaction participants,
it unfortunately impacts how long it will take the data to become consistent for the
atomic business transaction (see eventual consistency later in this section).
Soft state (the S part of BASE) describes the situation where a distributed transaction
is in progress and the state of the atomic business request is not yet complete (or in
some cases not even known). In the customer registration example shown in Figure
9-12, soft state occurs when the customer profile information is inserted (and com‐
mitted) in the Profile table, but the support contract and billing information are not.
The unknown part of soft state can occur if, using the same example, all three services
work in parallel to insert their corresponding data—the exact state of the atomic busi‐
ness request is not known at any point in time until all three services report back that
the data has been successfully processed. In the case of a workflow using asynchro‐
nous communication (see Chapter 11), the in-progress or final state of the distributed
transaction is usually difficult to determine.
Eventual consistency (the E part of BASE) means that given enough time, all parts of
the distributed transaction will complete successfully and all of the data is in sync
with one another. The type of eventual consistency pattern used and the way errors
are handled dictates how long it will take for all of the data sources involved in the
distributed transaction to become consistent.
The next section describes the three types of eventual consistency patterns and the
corresponding trade-offs associated with each pattern.
Eventual Consistency Patterns
Distributed architectures rely heavily on eventual consistency as a trade-off for better
operational architecture characteristics such as performance, scalability, elasticity,
fault tolerance, and availability. While there are numerous ways to achieve eventual
consistency between data sources and systems, the three main patterns in use today
are the background synchronization pattern, orchestrated request-based pattern, and
the event-based pattern.
To better describe each pattern and illustrate how they work, consider again the cus‐
tomer registration process from the Sysops Squad application we discussed earlier in
Eventual Consistency Patterns
|
267
Figure 9-13. In this example, three separate services are involved in the customer reg‐
istration process: a Customer Profile Service that maintains basic profile information,
a Support Contract Service that maintains products covered under the Sysops Squad
repair plan for each customer, and a Billing Payment Service that charges the cus‐
tomer for the support plan. Notice in the figure that customer 123 is a subscriber to
the Sysops Squad service, and therefore has data in each of the corresponding tables
owned by each service.
Figure 9-13. Customer 123 is a subscriber in the Sysops Squad application
Customer 123 decides they are no longer interested in the Sysops Squad support plan,
so they unsubscribe from the service. As shown in Figure 9-14, the Customer Profile
Service receives this request from the user interface, removes the customer from the
Profile table, and returns a confirmation to the customer that they are successfully
unsubscribed and will no longer be billed. However, data for that customer still exists
in the Contract table owned by the Support Contract Service and the Billing table
owned by the Billing Payment Service.
Figure 9-14. Data is out of sync after the customer unsubscribes from the support plan
268
|
Chapter 9: Data Ownership and Distributed Transactions
We will use this scenario to describe each of the eventual consistency patterns for get‐
ting all of the data in sync for this atomic business request.
Background Synchronization Pattern
The background synchronization pattern uses a separate external service or process to
periodically check data sources and keep them in sync with one another. The length
of time for data sources to become eventually consistent using this pattern can vary
based on whether the background process is implemented as a batch job running
sometime in the middle of the night, or a service that wakes up periodically (say,
every hour) to check the consistency of the data sources.
Regardless of how the background process is implemented (nightly batch or peri‐
odic), this pattern usually has the longest length of time for data sources to become
consistent. However, in many cases data sources do not need to be kept in sync
immediately. Consider the customer unsubscribe example in Figure 9-14. Once a cus‐
tomer unsubscribes, it really doesn’t matter that the support contract and billing
information for that customer still exists. In this case, eventual consistency done dur‐
ing the night is a sufficient amount of time to get the data in sync.
One of the challenges of this pattern is that the background process used to keep all
the data in sync must know what data has changed. This can be done through an
event stream, a database trigger, or reading data from source tables and aligning tar‐
get tables with the source data. Regardless of the technique used to identify changes,
the background process must have knowledge of all the tables and data sources
involved in the transaction.
Figure 9-15 illustrates the use of the background synchronization pattern for the
Sysops Squad unregister example. Notice that at 11:23:00 the customer issues a
request to unsubscribe from the support plan. The Customer Profile Service receives
the request, removes the data, and one second later (11:23:01) responds back to the
customer that they have been successfully unsubscribed from the system. Then, at
23:00 the background batch synchronization process starts. The background syn‐
chronization process detects that customer 123 has been removed either through
event streaming or primary table versus secondary table deltas, and deletes the data
from the Contract and Billing tables.
This pattern is good for overall responsiveness because the end user doesn’t have to
wait for the entire business transaction to complete (in this case, unsubscribing from
the support plan). But, unfortunately, some serious trade-offs with this eventual con‐
sistency pattern.
Eventual Consistency Patterns
|
269
Figure 9-15. The background synchronization pattern uses an external process to ensure
data consistency
The biggest disadvantage of the background synchronization pattern is that it couples
all of the data sources together, thus breaking every bounded context between the
data and the services. Notice in Figure 9-16 that the background batch synchroniza‐
tion process must have write access to each of the tables owned by the corresponding
services, meaning that all of the tables effectively have shared ownership between the
services and the background synchronization process.
This shared data ownership between the services and the background synchroniza‐
tion process is riddled with issues, and emphasizes the need for tight bounded con‐
texts within a distributed architecture. Structural changes made to the tables owned
by each service (changing a column name, dropping a column, and so on) must also
be coordinated with an external background process, making changes difficult and
time-consuming.
In addition to difficulties with change control, problems occur with regard to dupli‐
cated business logic as. In looking at Figure 9-15, it might seem fairly straightforward
that the background process would simply perform a DELETE operation on all rows in
the Contract and Billing tables containing customer 123. However, certain business
rules may exist within these services for the particular operation.
270
|
Chapter 9: Data Ownership and Distributed Transactions
Figure 9-16. The background synchronization pattern is coupled to the data sources,
therefore breaking the bounded context and data ownership
For example, when a customer unsubscribes, their existing support contracts and bill‐
ing history are kept for three months in the event the customer decides to resubscribe
to the support plan. Therefore, rather than deleting the rows in those tables, a
remove_date column is set with a long value representing the date the rows should be
removed (a zero value in this column indicates an active customer). Both services
check the remove_date daily to determine which rows should be removed from their
respective tables. The question is, where is that business logic located? The answer, of
course, is in the Support Contract and Billing Payment Services—oh, and also the
background batch process!
The background synchronization eventual consistency pattern is not suitable for dis‐
tributed architectures requiring tight bounded contexts (such as microservices)
where the coupling between data ownership and functionality is a critical part of the
architecture. Situations where this pattern is useful are closed (self-contained) hetero‐
geneous systems that don’t communicate with each other or share data.
For example, consider a contractor order entry system that accepts orders for build‐
ing materials, and another separate system (implemented in a different platform) that
does contractor invoicing. Once a contractor orders supplies, a background synchro‐
nization process moves those orders to the invoicing system to generate invoices.
When a contractor changes an order or cancels it, the background synchronization
process moves those changes to the invoicing system to update the invoices. This is a
good example of systems becoming eventually consistent, with the contractor order
always in sync between the two systems.
Eventual Consistency Patterns
|
271
Table 9-5 summarizes the trade-offs for the background synchronization pattern for
eventual consistency.
Trade-Offs
Table 9-5. Background synchronization pattern trade-offs
Advantages
Disadvantages
Services are decoupled
Good responsiveness
Data source coupling
Complex implementation
Breaks bounded contexts
Business logic may be duplicated
Slow eventual consistency
Orchestrated Request-Based Pattern
A common approach for managing distributed transactions is to make sure all of the
data sources are synchronized during the course of the business request (in other
words, while the end user is waiting). This approach is implemented through what is
known as the orchestrated request-based pattern.
Unlike the previous background synchronization pattern or the event-based pattern
described in the next section, the orchestrated request-based pattern attempts to pro‐
cess the entire distributed transaction during the business request, and therefore
requires some sort of orchestrator to manage the distributed transaction. The orches‐
trator, which can be a designated existing service or a new separate service, is respon‐
sible for managing all of the work needed to process the request, including
knowledge of the business process, knowledge of the participants involved, multicast‐
ing logic, error handling, and contract ownership.
One way to implement this pattern is to designate one of the primary services
(assuming there is one) to manage the distributed transaction. This technique, illus‐
trated in Figure 9-17, designates one of the services to take on the role as orchestrator
in addition to its other responsibilities, which in this case is the Customer Profile
Service.
272
| Chapter 9: Data Ownership and Distributed Transactions
Figure 9-17. The Customer Profile Service takes on the role of an orchestrator for the dis‐
tributed transaction
Although this approach avoids the need for a separate orchestration service, it tends
to overload the responsibilities of the service designated as the distributed transaction
orchestrator. In addition to the role of an orchestrator, the designated service manag‐
ing the distributed transaction must perform its own responsibilities as well. Another
drawback to this approach is that it lends itself to tight coupling and synchronous
dependencies between services.
The approach we generally prefer when using the orchestrated request-based pattern
is to use a dedicated orchestration service for the business request. This approach,
illustrated in Figure 9-18, frees up the Customer Profile Service from the responsibil‐
ity of managing the distributed transaction and places that responsibility on a sepa‐
rate orchestration service.
We will use this separate orchestration service approach to describe how this eventual
consistency pattern works and the corresponding trade-offs with this pattern.
Eventual Consistency Patterns
|
273
Figure 9-18. A dedicated orchestration service takes on the role of an orchestrator for the
distributed transaction
Notice that at 11:23:00 the customer issues a request to unsubscribe from the Sysops
Squad support plan. The request is received by the Unsubscribe Orchestrator Service,
which then forwards the request synchronously to the Customer Profile Service to
remove the customer from the Profile table. One second later, the Customer Profile
Service sends back an acknowledgment to the Unsubscribe Orchestrator Service,
which then sends parallel requests (either through threads or some sort of asynchro‐
nous protocol) to both the Support Contract and Billing Payment Services. Both of
these services process the unsubscribe request, and then send an acknowledgment
back one second later to the Unsubscribe Orchestrator Service indicating they are
done processing the request. Now that all data is in sync, the Unsubscribe Orchestra‐
tor Service responds back to the client at 11:23:02 (two seconds after the initial
request was made), letting the customer know they were successfully unsubscribed.
The first trade-off to observe is that the orchestration approach generally favors data
consistency over responsiveness. Adding a dedicated orchestration service not only
adds additional network hops and service calls, but depending on whether the
orchestrator executes calls serially or in parallel, additional time is needed for the
back-and-forth communication between the orchestrator and the services it’s calling.
274
|
Chapter 9: Data Ownership and Distributed Transactions
Response time could be improved in Figure 9-18 by executing the Customer Profile
request at the same time as the other services, but we chose to do that operation syn‐
chronously for error handling and consistency reasons. For example, if the customer
could not be deleted from the Profile table because of an outstanding billing charge,
no other action is needed to reverse the operations in the Support Contract and Bill‐
ing Payment Services. This represents another example of consistency over
responsiveness.
Besides responsiveness, the other trade-off with this pattern is complex error han‐
dling. While the orchestrated request-based pattern might seem straightforward, con‐
sider what happens when the customer is removed from the Profile table and
Contract table, but an error occurs when trying to remove the billing information
from the Billing table, as illustrated in Figure 9-19. Since the Profile and Support
Contract Services individually committed their operations, the Unsubscribe Orches‐
trator Service must now decide what action to take while the customer is waiting for
the request to be processed:
1. Should the orchestrator send the request again to the Billing Payment Service for
another try?
2. Should the orchestrator perform a compensating transaction and have the Sup‐
port Contract and Customer Profile Services reverse their update operations?
3. Should the orchestrator respond to the customer that an error occurred and to
wait a bit before trying again, while trying to repair the inconsistency?
4. Should the orchestrator ignore the error in hopes that some other process will
deal with the issue and respond to the customer that they have been successfully
unsubscribed?
This real-world scenario creates a messy situation for the orchestrator. Because this is
the eventual consistency pattern used, there is no other means to correct the data and
get things back in sync (therefore negating options 3 and 4 in the preceding list). In
this case, the only real option for the orchestrator is to try to reverse the distributed
transaction—in other words, issue a compensating update to reinsert the customer in
the Profile table and set the remove_date column in the Contract table back to zero.
This would require the orchestrator to have all of the necessary information to rein‐
sert the customer, and that no side effects occur when creating a new customer (such
as initializing the billing information or support contracts).
Eventual Consistency Patterns
|
275
Figure 9-19. Error conditions are very hard to address when using the orchestrated
request-based pattern
Another complication with compensating transactions in a distributed architecture is
failures that occur during compensation. For example, suppose a compensating trans‐
action was issued to the Customer Profile Service to reinsert the customer, and that
operation failed. Now what? Now the data is really out of sync, and there’s no other
service or process around to repair the problem. Most cases like these typically
require human intervention to repair the data sources and get them back in sync. We
go into more details about compensating transactions and transactional sagas in
“Transactional Saga Patterns” on page 324.
276
|
Chapter 9: Data Ownership and Distributed Transactions
Table 9-6 summarizes the trade-offs for the orchestrated request-based pattern for
eventual consistency.
Trade-Offs
Table 9-6. Orchestrated request-based pattern trade-offs
Advantages
Disadvantages
Services are decoupled
Timeliness of data consistency
Atomic business request
Slower responsiveness
Complex error handling
Usually requires compensating transactions
Event-Based Pattern
The event-based pattern is one of the most popular and reliable eventual consistency
patterns for most modern distributed architectures, including microservices and
event-driven architectures. With this pattern, events are used in conjunction with an
asynchronous publish-and-subscribe (pub/sub) messaging model to post events
(such as customer unsubscribed) or command messages (such as unsubscribe cus
tomer) to a topic or event stream. Services involved in the distributed transaction lis‐
ten for certain events and respond to those events.
The eventual consistency time is usually short for achieving data consistency because
of the parallel and decoupled nature of the asynchronous message processing. Serv‐
ices are highly decoupled from one another with this pattern, and responsiveness is
good because the service triggering the eventual consistency event doesn’t have to
wait for the data synchronization to occur before returning information to the
customer.
Figure 9-20 illustrates how the event-based pattern for eventual consistency works.
Notice that the customer issues the unsubscribe request to the Customer Profile Ser‐
vice at 11:23:00. The Customer Profile Service receives the request, removes the cus‐
tomer from the Profile table, publishes a message to a message topic or event stream,
and returns information one second later letting the customer know they were suc‐
cessfully unsubscribed. At around the same time this happens, both the Support Con‐
tract and Billing Payment Services receive the unsubscribe event and perform
whatever functionality is needed to unsubscribe the customer, making all the data
sources eventually consistent.
Eventual Consistency Patterns
|
277
Figure 9-20. The event-based pattern uses asynchronous publish-and-subscribe messag‐
ing or event streams to achieve eventual consistency
For implementations using standard topic-based publish-and-subscribe messaging
(such as ActiveMQ, RabbitMQ, AmazonMQ, and so on), services responding to the
event must be set up as durable subscribers to ensure no messages are lost if the mes‐
sage broker or the service receiving the message fails. A durable subscriber is similar
in concept to persistent queues in that the subscriber (in this case, the Support Con‐
tract Service and Billing Payment Service) does not need to be available at the time
the message is published, and subscribers are guaranteed to receive the message once
they become available. In the case of event streaming implementations, the message
broker (such as Apache Kafka) must always persist the message and make sure it is
available in the topic for a reasonable amount of time.
The advantages of the event-based pattern are responsiveness, timeliness of data con‐
sistency, and service decoupling. However, similar to all eventual consistency pat‐
terns, the main trade-off of this pattern is error handling. If one of the services (for
example, the Billing Payment Service illustrated in Figure 9-20) is not available, the
fact that it is a durable subscriber means that eventually it will receive and process the
event when it does become available. However, if the service is processing the event
and fails, things get complicated quickly.
278
| Chapter 9: Data Ownership and Distributed Transactions
Most message brokers will try a certain number of times to deliver a message, and
after repeated failures by the receiver, the broker will send the message to a dead letter
queue (DLQ). This is a configurable destination where the event is stored until an
automated process reads the message and tries to fix the problem. If it can’t be
repaired programmatically, the message is then typically sent to a human for manual
processing.
Table 9-7 lists the trade-offs for the event-based pattern for eventual consistency.
Trade-Offs
Table 9-7. Event-based pattern trade-offs
Advantages
Disadvantages
Services are decoupled
Timeliness of data consistency
Fast responsiveness
Complex error handling
Sysops Squad Saga: Data Ownership for Ticket Processing
Tuesday, January 18, 09:14
After talking with Dana and learning about data ownership and distributed transaction
management, Sydney and Addison quickly realized that breaking apart data and assigning data ownership to form tight bounded contexts wasn’t possible without both teams
collaborating on the solution.
“No wonder nothing ever seems to work around here,” observed Sydney. “We’ve always
had issues and arguments between us and the database team, and now I see the results
of our company treating us as two separate teams.”
“Exactly,” said Addison. “I’m glad we are working more closely with the data team now. So, from what
Dana said, the service that performs write actions on the data table owns the table, regardless of
what other services need to access the data in a read-only manner. In that case, looks like the User
Maintenance Service needs to own the data.”
Sydney agreed, and Addison created a general architecture decision record describing what to do
for single-table ownership scenarios:
Sysops Squad Saga: Data Ownership for Ticket Processing
|
279
ADR: Single Table Ownership for Bounded Contexts
Context
When forming bounded contexts between services and data, tables must be assigned
ownership to a particular service or group of services.
Decision
When only one service writes to a table, that table will be assigned ownership to that service. Furthermore, services requiring read-only access to a table in another bounded context cannot directly access the database or schema containing that table.
Per the database team, table ownership is defined as the service that performs write operations on a table. Therefore, for single table ownership scenarios, regardless of how many
other services need to access the table, only one service is ever assigned an owner, and
that owner is the service that maintains the data.
Consequences
Depending on the technique used, services requiring read-only access to a table in
another bounded context may incur performance and fault-tolerance issues when accessing data in a different bounded context.
Now that Sydney and Addison better understood table ownership and how to form bounded contexts between the service and the data, they started to work on the survey functionality. The Ticket
Completion Service would write the timestamp the ticket was completed and the expert who performed the job to the survey table. The Survey Service would write the timestamp the survey was
sent to the customer, and also insert all of the survey results once the survey is received.
“This isn’t so hard now that I better understand bounded contexts and table ownership,” said
Sydney.
“OK, let’s move on to the survey functionality,” said Addison.
“Oops,” said Sydney. “Both the Ticket Completion Service and the Survey Service write to the Survey
table.”
“That’s what Dana called joint-table ownership,” said Addison.
“So, what are our options?” asked Sydney.
“Since splitting up the table won’t work, it really leaves us with only two options,” said Addison. “We
can use a common data domain so that both services own the data, or we can use the delegate
technique and assign only one service as the owner.”
“I like the common data domain. Let both services write to the table and share a common schema,”
said Sydney.
“Except that won’t work in this scenario,” said Addison. “The Ticket Completion Service is already talking to the common ticketing data domain. Remember, a service can’t connect to multiple schemas.”
280
|
Chapter 9: Data Ownership and Distributed Transactions
“Oh, right,” said Sydney. “Wait, I know, just add the survey tables to the ticketing data domain
schema.”
“But now we are starting to combine all the tables back together.” said Addison. “Pretty soon we’ll be
right back to a monolithic database again.”
“So what do we do?” asked Sydney.
“Wait, I think I see a good solution here,” said Addison. “You know how the Ticket Completion Service
has to send a message to the Survey Service anyway to kick off the survey process once a ticket is
complete? What if we passed in the necessary data along with that message so that the Survey Service can insert the data when it creates the customer survey?”
“That’s brilliant,” said Sydney. “That way, the Ticket Completion doesn’t need any access to the Survey
table.”
Addison and Sydney agreed that the Survey Service would own the Survey table, and would use the
delegation technique to pass data when the table notifies the Survey Service to kick off the survey
process as illustrated in Figure 9-21. Addison wrote an architecture decision record for this decision.
Figure 9-21. Survey Service owns the data using the delegation technique
ADR: Survey Service Owns the Survey Table
Context
Both the Ticket Completion Service and the Survey Service write to the Survey table.
Because this is a joint ownership scenario, the alternatives are to use a common shared
data domain or use the delegation technique. Table splitting is not an option because of
the structure of the Survey table.
Decision
The Survey Service will be the single owner of the Survey table, meaning it is the only service that can perform write operations to that table.
Once a ticket is marked as complete and is accepted by the system, the Ticket Completion
Service needs to send a message to the Survey Service to kick off the customer survey
processing. Since the Ticket Completion Service is already sending a notification event,
Sysops Squad Saga: Data Ownership for Ticket Processing
|
281
the necessary ticket information can be passed along with that event, thus eliminating the
need for the Ticket Completion Service to have any access to the Survey table.
Consequences
All of the necessary data that the Ticket Completion Service needs to insert into the Survey table will need to be sent as part of the payload when triggering the customer survey
process.
In the monolithic system, the ticket completion inserted the survey record as part of the
completion process. With this decision, the creation of the survey record is a separate
activity from the ticket creation process and is now handled by the Survey Service.
282
|
Chapter 9: Data Ownership and Distributed Transactions
CHAPTER 10
Distributed Data Access
Monday, January 3, 12:43
“Now that we’ve assigned ownership of the expert profile table to the User Management
Service,” said Sydney, “how should the Ticket Assignment Service get to the expert location and skills data? As I said before, with the number of reads it does to the database, it’s
really not feasible to make a remote call every time it needs to query the table.”
“Can you modify the way the assignment algorithm works so that we can reduce the
number of queries it needs?” asked Addison.
“Beats me,” replied Sydney. “Taylen’s the one who usually maintains those algorithms.”
Addison and Sydney met with Taylen to discuss the data access issue and to see if Taylen could
modify the expert assignment algorithms to reduce the nimber of database calls to the expert profile table.
“Are you kidding me?” asked Taylen. “There’s no way I can rewrite the assignment algorithms to do
what you are asking. Absolutely no way at all.”
“But our only other option is to make remote calls to the User Management Service every time the
assignment algorithm needs expert data,” said Addison.
“What?” screamed Taylen. “We can’t do that!”
“That what I said as well,” said Sydney. “That means we are back to square one again. This distributed
architecture stuff is hard. I hate to say this, but I am actually starting to miss the monolithic application. Wait, I know. What if we made messaging calls to the User Maintenance Service instead of
using REST?”
283
“That’s the same thing,” said Taylen. “I still have to wait for the information to come back, whether we
use messaging, REST, or any other remote access protocol. That table simply needs to be in the same
data domain as the ticketing tables.”
“There’s got to be another solution to access data we no longer own,” said Addison. “Let me check
with Logan.”
In most monolithic systems using a single database, developers don’t give a second
thought to reading database tables. SQL table joins are commonplace, and with a sim‐
ple query all necessary data can be retrieved in a single database call. However, when
data is broken into separate databases or schemas owned by distinct services, data
access for read operations starts to become hard.
This chapter describes the various ways services can gain read access to data they
don’t own—in other words, outside the bounded context of the services needing the
data. The four patterns of data access we discuss in this chapter include the Interservice Communication pattern, Column Schema Replication pattern, Replicated
Cache pattern, and the Data Domain pattern.
Each of these data access patterns has its share of advantages and disadvantages. Yes,
once again, trade-offs. To better describe each of these patterns, we will return to our
Wishlist Service and a Catalog Service example from Chapter 9. The Wishlist Service
shown in Figure 10-1 maintains a list of items a customer may want to eventually
purchase, and contains the customer ID, item ID, and date the item was added in the
corresponding Wishlist table. The Catalog Service is responsible for maintaining all
of the items the company sells, and includes the item ID, item description, and static
product dimension information, such as the weight, height, length, and so on.
In this example, when a request is made from a customer to display in their wish list,
both the item ID and and the item description (item_desc) are returned to the cus‐
tomer. However, the Wishlist Service does not have the item description in its table;
that data is owned by the Catalog Service in a tightly formed bounded context provid‐
ing change control and data ownership. Therefore, the architect must use one of the
data access patterns outlined in this chapter to ensure the Wishlist Service can obtain
the product descriptions from the Catalog Service.
284
|
Chapter 10: Distributed Data Access
Figure 10-1. Wishlist Service needs item descriptions but doesn’t have access to the
product table containing the data
Interservice Communication Pattern
The Interservice Communication pattern is by far the most common pattern for
accessing data in a distributed system. If one service (or system) needs to read data
that it cannot access directly, it simply asks the owning service or system for it by
using some sort of remote access protocol. What can be more simple?
As with most things in software architecture, all is not as it seems. While simple, this
common data access technique is unfortunately riddled with disadvantages. Consider
Figure 10-2: the Wishlist Service makes a synchronous remote access call to the Cata‐
log Service, passing in a list of item IDs in exchange for a list of corresponding item
descriptions.
Notice that for every request to get a customer wish list, the Wishlist Service must
make a remote call to the Catalog Service to get the item descriptions. The first issue
that occurs with this pattern is slower performance due to network latency, security
latency, and data latency. Network latency is the packet transmission time to and from
a service (usually somewhere between 30 ms and 300 ms). Security latency occurs
when the endpoint to the target service requires additional authorization to perform
the request. Security latency can vary greatly depending on the level of security on the
endpoint being accessed, but could be anywhere between 20 ms and 400 ms for most
systems. Data latency describes the situation where multiple database calls need to be
made to retrieve the necessary information to pass back to the end user. In this case,
rather than a single SQL table join statement, an additional database call must be
made by the Catalog Service to retrieve the item description. This might add
Interservice Communication Pattern
|
285
anywhere from 10 ms to 50 ms additional processing time. Add all of that up, and the
latency could be up to one second just to get the item descriptions.
Figure 10-2. Interservice communication data access pattern
Another big disadvantage of this pattern is service coupling. Because the Wishlist
must rely on the Catalog Service being available, the services are therefore both
semantically and statically coupled, meaning that if the Catalog Service is not avail‐
able, neither is the Wishlist Service. Furthermore, because of the tight static coupling
between the Wishlist Service and the Catalog Service, as the Wishlist Service scales to
meet additional demand volume, so must the Catalog Service.
Table 10-1 summarizes the trade-offs associated with the interservice communication
data access pattern.
Trade-Offs
Table 10-1. Trade-offs for the Interservice Communication data access pattern
286
Advantages
Disadvantages
Simplicity
No data volume issues
Network, data, and security latency (performance)
Scalability and throughput issues
No fault tolerance (availability issues)
Requires contracts between services
| Chapter 10: Distributed Data Access
Column Schema Replication Pattern
With the Column Schema Replication pattern, columns are replicated across tables,
therefore replicating the data and making it available to other bounded contexts. As
shown in Figure 10-3, the item_desc column is added to the Wishlist table, making
that data available to the Wishlist Service without having to ask the Catalog Service
for the data.
Figure 10-3. With the Column Schema Replication data access pattern, data is replicated
to other tables
Data synchronization and data consistency are the two biggest issues associated with
the Column Schema Replication data access pattern. Whenever a product is created,
removed from the catalog, or a product description changed, the Catalog Service
must somehow let the Wishlist Service (and any other services replicating the data)
know about the change. This is usually done through asynchronous communications
using queues, topics, or event streaming. Unless immediate transactional synchroni‐
zation is required, asynchronous communication is a preferred choice over synchro‐
nous communication because it increases responsiveness and reduces the availability
dependency between the services.
Another challenge with this pattern is that it is sometimes difficult to govern data
ownership. Because the data is replicated in tables belonging to other services, those
services can update the data, even though they don’t officially own the data. This in
turn creates even more data consistency issues.
Column Schema Replication Pattern
|
287
Even though the services are still coupled because of data synchronization, the service
requiring read access has immediate access to the data, and can do simple SQL joins
or queries to its own table to get the data. This increases performance, fault tolerance,
and scalability, all things that were disadvantages with the interservice communica‐
tion pattern.
While in general we caution against use of this data access pattern for scenarios such
as the Wishlist Service and Catalog Service example, some situations where it might
be a consideration are data aggregation, reporting, or situations where the other data
access patterns are not a good fit because of large data volumes, high responsiveness
requirements, or high-fault tolerance requirements.
Table 10-2 summarizes the trade-offs associated with the Column Schema Replica‐
tion data access pattern.
Trade-Offs
Table 10-2. Trade-offs for the Column Schema Replication data access pattern
Advantages
Disadvantages
Good data access performance
No scalability and throughput issues
No fault-tolerance issues
No service dependencies
Data consistency issues
Data ownership issues
Data synchronization is required
Replicated Caching Pattern
Most developers and architects think of caching as a technique for increasing overall
responsiveness. By storing data within an in-memory cache, retrieving data goes
from dozens of milliseconds to only a couple of nanoseconds. However, caching can
also be an effective tool for distributed data access and sharing. This pattern leverages
replicated in-memory caching so that data needed by other services is made available
to each service without them having to ask for it. A replicated cache differs from
other caching models in that data is held in-memory within each service and is con‐
tinuously synchronized so that all services have the same exact data at all times.
288
|
Chapter 10: Distributed Data Access
To better understand the replicated caching model, it’s useful to compare it to other
caching models to see the differences between them. The single in-memory caching
model is the simplest form of caching, where each service has its own internal inmemory cache. With this caching model (illustrated in Figure 10-4), in-memory data
is not synchronized between the caches, meaning each service has its own unique
data specific to that service. While this caching model does help increase responsive‐
ness and scalability within each service, it’s not useful for sharing data between serv‐
ices because of the lack of cache synchronization between the services.
Figure 10-4. With a single in-memory cache, each service contains its own unique data
The other caching model used in distributed architectures is distributed caching. As
illustrated in Figure 10-5, with this caching model, data is not held in-memory within
each service, but rather held externally within a caching server. Services, using a pro‐
prietary protocol, make requests to the caching server to retrieve or update shared
data. Note that unlike the single in-memory caching model, data can be shared
among the services.
The distributed cache model is not an effective caching model to use for the replica‐
ted caching data access pattern for several reasons. First, there’s no benefit to the
fault-tolerance issues found with the Interservice Communication pattern. Rather
than depending on a service to retrieve data, the dependency has merely shifted to
the caching server.
Because the cache data is centralized and shared, the distributed cache model allows
other services to update data, thereby breaking the bounded context regarding data
ownership. This can cause data inconsistencies between the cache and the owning
database. While this can sometimes be addressed through strict governance, it is nev‐
ertheless an issue with this caching model.
Lastly, since access to the centralized distributed cache is through a remote call, net‐
work latency adds additional retrieval time for the data, thus impacting overall
responsiveness as compared to an in-memory replicated cache.
Replicated Caching Pattern
|
289
Figure 10-5. A distributed cache is external from the services
With replicated caching, each service has its own in-memory data that is kept in sync
between the services, allowing the same data to be shared across multiple services.
Notice in Figure 10-6 that there is no external cache dependency. Each cache instance
communicates with another so that when an update is made to a cache, that update is
immediately (behind the scenes) asynchronously propagated to other services using
the same cache.
Figure 10-6. With a replicated cache, each service contains the same in-memory data
Not all caching products support replicated caching, so it’s important to check with
the caching product vendor to ensure support for the replicated caching model. Some
of the popular products that do support replicated caching include Hazelcast, Apache
Ignite, and Oracle Coherence.
To see how replicated caching can address distributed data access, we’ll return to our
Wishlist Service and Catalog Service example. In Figure 10-7, the Catalog Service
owns an in-memory cache of product descriptions (meaning it is the only service that
can modify the cache), and the Wishlist Service contains a read-only in-memory rep‐
lica of the same cache.
290
|
Chapter 10: Distributed Data Access
Figure 10-7. Replicated caching data access pattern
With this pattern, the Wishlist Service no longer needs to make calls to the Catalog
Service to retrieve product descriptions—they’re already in-memory within the
Wishlist Service. When updates are made to the product description by the Catalog
Service, the caching product will update the cache in the Wishlist Service to make the
data consistent.
The clear advantages of the replicated caching pattern are responsiveness, fault toler‐
ance, and scalability. Because no explicit interservice communication is required
between the services, data is readily available in-memory, providing the fastest possi‐
ble access to data a service doesn’t own. Fault tolerance is also well supported with
this pattern. Even if the Catalog Service goes down, the Wishlist Service can continue
to operate. Once the Catalog Service comes back up, the caches connect to one
another without any disruption to the Wishlist Service. Lastly, with this pattern, the
Wishlist Service can scale independently from the Catalog Service.
With all these clear advantages, how could there possibly be a trade-off with this pat‐
tern? As the first law of software architecture states in our book, The Fundamentals of
Software Architecture, everything in software architecture is a trade-off, and if an
architect thinks they have discovered something that isn’t a trade-off, it means they
just haven’t identified the trade-off yet.
Replicated Caching Pattern
|
291
The first trade-off with this pattern is a service dependency with regard to the cache
data and startup timing. Since the Catalog Service owns the cache and is responsible
for populating the cache, it must be running when the initial Wishlist Service starts
up. If the Catalog Service is unavailable, the initial Wishlist Service must go into a
wait state until a connection with the Catalog Service is established. Notice that only
the initial Wishlist Service instance is impacted by this startup dependency; if the
Catalog Service is down, other Wishlist instances can be started up, with the cache
data transferred from one of the other Wishlist instances. It’s also important to note
that once the Wishlist Service starts and has the data in the cache, it is not necessary
for the Catalog Service to be available. Once the cache is made available in the Wish‐
list Service, the Catalog Service can come up and down without impacting the Wish‐
list Service (or any of its instances).
The second trade-off with this pattern is that of data volumes. If the volume of data is
too high (such as exceeding 500 MB), the feasibility of this pattern diminishes
quickly, particularly with regard to multiple instances of services needing the data.
Each service instance has its own replicated cache, meaning that if the cache size of
500 MB and 5 instances of a service are required, the total memory used is 2.5 GB.
Architects must analyze both the size of the cache and the total number of services
instances needing the cached data to determine the total memory requirements for
the replicated cache.
A third trade-off is that the replicated caching model usually cannot keep the data
fully in sync between services if the rate of change of the data (update rate) is too
high. This varies based on the size of the data and the replication latency, but in gen‐
eral this pattern is not well suited for highly volatile data (such as product inventory
counts). However, for relatively static data (such as a product description), this pat‐
tern works well.
The last trade-off associated with this pattern is that of configuration and setup man‐
agement. Services know about each other in the replicated caching model through
TCP/IP broadcasts and lookups. If the TCI/IP broadcast and lookup range is too
broad, it can take a long time to establish the socket-level handshake between serv‐
ices. Cloud-based and containerized environments make this particularly challenging
because of the lack of control over IP addresses and the dynamic nature of IP
addresses associated with these environments.
292
|
Chapter 10: Distributed Data Access
Table 10-3 lists the trade-offs associated with the replicated cache data access pattern.
Trade-Offs
Table 10-3. Trade-offs associated with the replicated caching data access pattern
Advantages
Disadvantages
Good data access performance
No scalability and throughput issues
Good level of fault tolerance
Data remains consistent
Data ownership is preserved
Cloud and containerized configuration can be hard
Not good for high data volumes
Not good for high update rates
Initial service startup dependency
Data Domain Pattern
In the previous chapter, we discussed the use of a data domain to resolve joint owner‐
ship, where multiple services both need to write data to the same table. Tables that are
shared between services are put into a single schema that is then shared by both serv‐
ices. That same pattern can be used for data access as well.
Consider the Wishlist Service and Catalog Service problem again, where the Wishlist
Service needs access to the product descriptions but does not have access to the table
containing those descriptions. Suppose the Interservice Communication pattern is
not a feasible solution because of reliability issues with the Catalog Service as well as
the performance issues with network latency and the additional data retrieval. Also
suppose using the Column Schema Replication pattern is not feasible because of the
need for high levels of data consistency. Finally, suppose that the Replicated Cache
pattern isn’t an option because of the high data volumes. The only other solution is to
create a data domain, combining the Wishlist and Product tables in the same shared
schema, accessible to both the Wishlist Service and the Catalog Service.
Figure 10-8 illustrates the use of this data access pattern. Notice that the Wishlist and
Product tables are no longer owned by either service, but rather shared between
them, forming a broader bounded context. With this pattern, gaining access to the
product descriptions in the Wishlist Service is a matter of a simple SQL join between
the two tables.
Data Domain Pattern
|
293
Figure 10-8. Data domain data access pattern
While the sharing of data is generally discouraged in a distributed architecture, this
pattern has huge benefits over the other data access patterns. First of all, the services
are completely decoupled from each other, thereby resolving any availability depend‐
ency, responsiveness, throughput, and scalability issues. Responsiveness is very good
with this pattern because the data is available using a normal SQL call, removing the
need to do additional data aggregations within the functionality of the service (as is
required with the Replicated Cache pattern).
Both data consistency and data integrity rate very high with the Data Domain pat‐
tern. Since multiple services access the same data tables, data does not need to be
transferred, replicated, or synchronized. Data integrity is preserved in this pattern in
the sense that foreign-key constraints can now be enforced between the tables. In
addition, other database artifacts, such as views, stored procedures, and triggers, can
exist within the data domain. As a matter of fact, the preservation of these integrity
constraints and database artifacts is another driver for the use of the Data Domain
pattern.
With this pattern, no additional contracts are needed to transfer data between serv‐
ices—the table schema becomes the contract. While this is an advantage for this pat‐
tern, it’s a trade-off as well. The contracts used with the interservice communication
pattern and the Replicated Cache pattern form an abstraction layer over the table
schema, allowing changes to the table structures to remain within a tight bounded
context and not impact other services. However, this pattern forms a broader boun‐
ded context, requiring multiple services to possibly change when the structure to any
of the tables in the data domain changes.
294
| Chapter 10: Distributed Data Access
Another disadvantage of this pattern is that it can possibly open up security issues
associated with data access. For example, in Figure 10-8 the Wishlist Service has com‐
plete access to all the data within the data domain. While this is OK in the Wishlist
and Catalog Service example, there might be times when services accessing the data
domain shouldn’t have access to certain data. A tighter bounded context with strict
service ownership can prevent other services from accessing certain data through the
contracts used to pass the data back and forth.
Table 10-4 lists trade-offs associated with the data domain data access pattern.
Trade-Offs
Table 10-4. Trade-offs associated with the data domain data access pattern
Advantages
Disadvantages
Good data access performance
No scalability and throughput issues
No fault tolerance issues
No service dependency
Data remains consistent
Broader bounded context to manage data changes
Data ownership governance
Data access security
Sysops Squad Saga: Data Access for Ticket Assignment
Thursday, March 3, 14:59
Logan explained the various methods for data access within a distributed architecture,
and also outlined the corresponding trade-offs of each technique. Addison, Sydney, and
Taylen then had to come to a decision about which technique to use.
“Unless we start consolidating all of these services, I guess we are stuck with the fact that
the Ticket Assignment needs to somehow get to the expert profile data, and fast,” said
Taylen.
“OK,” said Addison. “So service consolidation is out because these services are in entirely different
domains, and the shared data domain option is out for the same reasons we talked about before—
we cannot have the Ticket Assignment Service connecting to two different databases.”
“So, that leaves us with one of two choices.” said Sydney. “Either we use interservice communication
or replicated caching.”
“Wait. Let’s explore the replicated caching option for a minute,” said Taylen. “How much data are we
talking about here?”
Sysops Squad Saga: Data Access for Ticket Assignment
|
295
“Well,” said Sydney, “we have 900 experts in the database. What data does the Ticket Assignment
Service need from the expert profile table?”
“It’s mostly static information as we get the current expert location feeds from elsewhere. So, it
would be the expert’s skill, their service location zones, and their standard scheduled availability,”
said Taylen.
“OK, so that’s about 1.3 KB of data per expert. And since we have 900 experts total, that would be…
about 1200 KB of data total. And the data is relatively static,” said Sydney.
“Hmm, that isn’t much data to store in memory,” said Taylen.
“Let’s not forget that if we used a replicated cache, we would have to take into account how many
instances we would have for the User Management Service as well as the Ticket Assignment Service,” said Addison. “Just to be on the safe side, we should use the maximum number of instances of
each we expect.”
“I’ve got that information,” said Taylen. “We expect to have only a maximum of two instances of the
User Management Service, and a maximum of four at our highest peak for the Ticket Assignment
Service.”
“That’s not much total in-memory data,” observed Sydney.
“No, it’s not,” said Addison. “OK, let’s analyze the trade-offs using the hypothesis-based approach we
tried earlier. I suggest that we should go with the in-memory replicated cache option to cache only
the data necessary for the Ticket Assignment Service. Any other trade-offs you can think of?”
Both Taylen and Sydney sat there for while trying to think of some negatives for the replicated cache
approach.
“What if the User Management Service goes down?” asked Sydney.
“As long as the cache is populated, then the Ticket Assignment Service would be fine,” said Addison.
“Wait, you mean to tell me that the data would be in-memory, even if the User Management Service
is unavailable?” asked Taylen.
“As long as the User Management Service starts before the Ticket Assignment Service, then yes,” said
Addison.
“Ah!” said Taylen. “Then there’s our first trade-off. Ticket assignment cannot function unless the User
Management Service is started. That’s not good.”
“But,” said Addison, “if we made remote calls to the User Management Service and it goes down, the
Ticket Assignment Service becomes nonoperational. At least with the replicated cache option, once
User Management is up and running, we are no longer dependent on it. So, replicated caching is
actually more fault tolerant in this case.”
“True,” said Taylen. “We just have to be careful about the startup dependency.”
296
|
Chapter 10: Distributed Data Access
“Anything else you can think of as a negative?” asked Addison, knowing another obvious trade-off
but wanting the development team to come up with it on their own.
“Um,” said Sydney, “yeah. I have one. What caching product are we going to use?”
“Ah,” said Addison, “that is in fact another trade-off. Have either of you done replicated caching
before? Or anyone on the development team for that matter?”
Both Taylen and Sydney shook their heads.
“Then we have some risk here,” said Addison.
“Actually,” said Taylen, “I’ve been hearing a lot about this caching technique for a while and have
been dying to try it out. I would volunteer to research some of the products and do some proof-ofconcepts on this approach.”
“Great,” said Addison. “In the meantime, I will research what the licensing cost would be for those
products as well, and if there’s any technical limitation with respect to our deployment environment.
You know, things like availability zone crossovers, firewalls, that sort of stuff.”
The team began their research and proof-of-concept work, and found that this is indeed not only a
feasible solution cost and effort wise, but would solve the issue of data access to the expert profile
table. Addison discussed this approach with Logan, who approved the solution. Addison created an
ADR outlining and justifying this decision.
ADR: Use of In-Memory Replicated Caching for Expert Profile Data
Context
The Ticket Assignment Service needs continuous access to the expert profile table, which
is owned by the User Management Service in a separate bounded context. Access to the
expert profile information can be done through interservice communication, in-memory
replicated caching, or a common data domain.
Decision
We will use replicated caching between the User Management Service and the Ticket
Assignment Service, with the User Management Service being the sole owner for write
operations.
Because the Ticket Assignment Service already connects to the shared ticket data domain
schema, it cannot connect to an additional schema. In addition, since the user management functionality and the core ticketing functionality are in two separate domains, we do
not want to combine the data tables in a single schema. Therefore, using a common data
domain is not an option.
Using an in-memory replicated cache resolves the performance and fault-tolerance issues
associated with the interservice communication option.
Sysops Squad Saga: Data Access for Ticket Assignment
|
297
Consequences
At least one instance of the User Management Service must be running when starting the
first instance of the Ticket Assignment Service.
Licensing costs for the caching product would be required for this option.
298
| Chapter 10: Distributed Data Access
CHAPTER 11
Managing Distributed Workflows
Tuesday, February 15, 14:34
Austen bolted into Logan’s office just after lunch. “I’ve been looking at the new architecture designs, and I want to help out. Do you need me to write up some ADRs or help
with some spikes? I’d be happy to write up the ADR that states that we’re only going to
use choreography in the new architecture to keep things decoupled.”
“Whoa, there, you maniac,” said Logan. “Where did you hear that? What gives you that
impression?”
“Well, I’ve been reading a lot about microservices, and everyone’s advice seems to be to keep things
highly decoupled. When I look at the patterns for communication, it seems that choreography is the
most decoupled, so we should always use it, right?”
"Always is a tricky term in software architecture. I had a mentor who had a memorable perspective
on this, who always said, Never use absolutes when talking about architecture, except when talking
about absolutes. In other words, never say never. I can’t think of many decisions in architecture where
always or never applies.”
“OK,” said Austen. “So how do architects decide between the different communication patterns?”
As part of our ongoing analysis of the trade-offs associated with modern distributed
architectures, we reach the dynamic part of quantum coupling, realizing many of the
patterns we described and named in Chapter 2. In fact, even our named patterns only
touch on the many permutations possible with modern architectures. Thus, an archi‐
tect should understand the forces at work so that they can make a most objective
trade-off analysis.
299
In Chapter 2, we identified three coupling forces when considering interaction mod‐
els in distributed architectures: communication, consistency, and coordination,
shown in Figure 11-1.
Figure 11-1. The dimensions of dynamic quantum coupling
In this chapter, we discuss coordination: combining two or more services in a dis‐
tributed architecture to form some domain-specific work, along with the many
attendant issues.
Two fundamental coordination patterns exist in distributed architectures: orchestra‐
tion and choreography. The fundamental topological differences between the two
styles is illustrated in Figure 11-2.
Orchestration is distinguished by the use of an orchestrator, whereas a choreographed
solution does not use one.
300
| Chapter 11: Managing Distributed Workflows
Figure 11-2. Orchestration versus choreography in distributed architectures
Orchestration Communication Style
The orchestration pattern uses an orchestrator (sometimes called a mediator) compo‐
nent to manage workflow state, optional behavior, error handling, notification, and a
host of other workflow maintenance. It is named for the distinguishing feature of a
musical orchestra, which utilizes a conductor to synchronize the incomplete parts of
the overall score to create a unified piece of music. Orchestration is illustrated in the
most generic representation in Figure 11-3.
In this example, services A-D are domain services, each responsible for its own boun‐
ded context, data, and behavior. The Orchestrator component generally doesn’t
include any domain behavior outside of the workflow it mediates. Notice that micro‐
services architectures have an orchestrator per workflow, not a global orchestrator
such as an enterprise service bus (ESB). One of the primary goals of the microservices
architecture style is decoupling, and using a global component such as an ESB creates
Orchestration Communication Style
|
301
an undesirable coupling point. Thus, microservices tend to have an orchestrator per
workflow.
Figure 11-3. Orchestration among distributed microservices
Orchestration is useful when an architect must model a complex workflow that
includes more than just the single “happy path,” but also alternate paths and error
conditions. However, to understand the basic shape of the pattern, we start with the
nonerror happy path. Consider a very simple example of Penultimate Electronics sell‐
ing a device to one of its customers online, shown in Figure 11-4.
This system passes the Place Order request to the Order Placement Orchestrator,
which makes a synchronous call to the Order Placement Service, which records the
order and returns a status message. Next, the mediator calls the Payment Service,
which updates payment information. Next, the orchestrator makes an asynchronous
call to the Fulfillment Service to handle the order. The call is asynchronous because
no strict timing dependencies exist for order fulfillment, unlike payment verification.
For example, if order fulfillment happens only a few times a day, there is no reason
for the overhead of a synchronous call. Similarly, the orchestrator then calls the Email
Service to notify the user of a successful electronics order.
If the world consisted of only happy paths, software architecture would be easy. How‐
ever, one of the primary hard parts of software architecture is error conditions and
pathways.
302
|
Chapter 11: Managing Distributed Workflows
Figure 11-4. A “happy path” workflow using an orchestrator to purchase electronic
equipment (note the asynchronous calls denoted by dotted lines for less time-sensitive
calls)
Consider two potential error scenarios for electronics purchasing. First, what hap‐
pens if the customer’s payment method is rejected? This error scenario appears in
Figure 11-5.
Here, the Order Placement Orchestrator updates the order via the Order Placement
Service as before. However, when trying to apply payment, it is rejected by the pay‐
ment service, perhaps because of an expired credit card number. In that case, the Pay‐
ment Service notifies the orchestrator, which then places a (typically) asynchronous
call to send a message to the Email Service to notify the customer of the failed order.
Additionally, the orchestrator updates the state of the Order Placement Service, which
still thinks this is an active order.
Notice in this example we’re allowing each service to maintain its own transactional
state, modeling our “Fairy Tale Saga(seo) Pattern” on page 333. One of the hardest parts
of modern architectures is managing transactions, which we cover in Chapter 12.
Orchestration Communication Style
|
303
Figure 11-5. Payment rejected error condition
In the second error scenario, the workflow has progressed further along: what hap‐
pens when the Fulfillment Service reports a back order? This error scenario appears
in Figure 11-6.
Figure 11-6. When an item is back-ordered, the orchestrator must rectify the state
304
|
Chapter 11: Managing Distributed Workflows
As you can see, the workflow preceeds as normal until the Fulfillment Service notifies
the orchestrator that the current item is out of stock, necessitating a back order. In
that case, the orchestrator must refund the payment (this is why many online services
don’t charge until shipment, not at order time) and update the state of the Order
Placement Service.
One interesting characteristic to note in Figure 11-6: even in the most elaborate error
scenarios, the architect wasn’t required to add additional communication paths that
weren’t already there to facilitate the normal workflow, which differs from the “Chor‐
eography Communication Style” on page 306.
General advantages of the orchestration communication style include the following:
Centralized workflow
As complexity goes up, having a unified component for state and behavior
becomes beneficial.
Error handling
Error handling is a major part of many domain workflows, assisted by having a
state owner for the workflow.
Recoverability
Because an orchestrator monitors the state of the workflow, an architect may add
logic to retry if one or more domain services suffers from a short-term outage.
State management
Having an orchestrator makes the state of the workflow queriable, providing a
place for other workflows and other transient states.
General disadvantages of the orchestration communication style include the
following:
Responsiveness
All communication must go through the mediator, creating a potential through‐
put bottleneck that can harm responsiveness.
Fault tolerance
While orchestration enhances recoverability for domain services, it creates a
potential single point of failure for the workflow, which can be addressed with
redundancy but adds more complexity.
Scalability
This communication style doesn’t scale as well as choreography because it has
more coordination points (the orchestrator), which cuts down on potential paral‐
lelism. As we discussed in Chapter 2, several dynamic coupling patterns utilize
choreography and thus achieve higher scale (notably “Time Travel Saga(sec) Pat‐
tern” on page 336 and “Anthology Saga(aec) Pattern” on page 349).
Orchestration Communication Style
|
305
Service coupling
Having a central orchestrator creates higher coupling between it and domain
components, which is sometimes necessary. The orchestration communication
style’s trade-offs appear in Table 11-1.
Trade-Offs
Table 11-1. Trade-offs for orchestration
Advantage
Disadvantage
Centralized workflow
Error handling
Recoverability
State management
Responsiveness
Fault tolerance
Scalability
Service coupling
Choreography Communication Style
Whereas the Orchestration Communication Style was named for the metaphorical
central coordination offered by an orchestrator, the choreography pattern visually
illustrates intent of the communication style that has no central coordination.
Instead, each service participates with the others, similar to dance partners. It isn’t an
ad hoc performance—the moves were planned beforehand by the choreographer/
architect but executed without a central coordinator.
Figure 11-4 described the orchestrated workflow for a customer purchasing electron‐
ics from Penultimate Electronics; the same workflow modeled in the choreography
communication style appears in Figure 11-7.
In this workflow, the initiating request goes to the first service in the chain of respon‐
sibility—in this case, the Order Placement Service. Once it has updated internal
records about the order, it sends an asynchronous request that the Payment Service
receives. Once payment has been applied, the Payment Service generates a message
received by the Fulfillment Service, which plans for delivery and sends a message to
the Email Service.
At first glance, the choreography solution seems simpler—fewer services (no orches‐
trator), and a simple chain of events/commands (messages). However, as with many
issues in software architecture, the difficulties lie not with the default paths but rather
with boundary and error conditions.
306
|
Chapter 11: Managing Distributed Workflows
Figure 11-7. Purchasing electronics using choreography
As in the previous section, we cover two potential error scenarios. The first results
from failed payment, as illustrated in Figure 11-8.
Figure 11-8. Error in payment in choreography
Rather than send a message intended for the Fulfillment Service, the Payment service
sends messages indicating failure to the Email Service and back to the Order Place‐
ment Service to update the order status. This alternate workflow doesn’t appear too
complex, with a single new communication link that didn’t exist before.
Choreography Communication Style
|
307
However, consider the increasing complexity imposed by the other error scenario for
a product back order, shown in Figure 11-9.
Figure 11-9. Managing the workflow error condition of product backlog
Many steps of this workflow have already completed before the event (out of stock)
that causes the error. Because each of these services implement its own transactional‐
ity (this is an example of the “Anthology Saga(aec) Pattern” on page 349), when an error
occurs, each service must issue compensating messages to other services. Once the
Fulfillment Service realizes the error condition, it should generate events suited to its
bounded context, perhaps a broadcast message subscribed to by the Email, Payment,
and Order Placement services.
The example shown in Figure 11-9 illustrates the dependency between complex
workflows and mediators. While the initial workflow in choreography illustrated in
Figure 11-7 seemed simpler than Figure 11-4, the error case (and others) keeps
adding more complexity to the choreographed solution. In Figure 11-10, each error
scenario forces domain services to interact with each other, adding communication
links that weren’t necessary for the happy path.
308
|
Chapter 11: Managing Distributed Workflows
Figure 11-10. Error conditions in choreography typically add communication links
Every workflow that architects need to model in software has a certain amount of
semantic coupling—the inherent coupling that exists in the problem domain. For
example, the process of assigning a ticket to a Sysops Squad member has a certain
workflow: a client must request service, skills must be matched to particular special‐
ists, then cross-referenced to schedules and locations. The way an architect models
that interaction is the implementation coupling.
The semantic coupling of a workflow is mandated by the domain requirements of the
solution and must be modeled somehow. However clever an architect is, they cannot
reduce the amount of semantic coupling, but their implementation choices may
increase it. This doesn’t mean that an architect might not push back on impractical or
impossible semantics defined by business users—some domain requirements create
extraordinarily difficult problems in architecture.
Here is a common example. Consider the standard layered monolithic architecture
compared to the more modern style of a modular monolith, shown in Figure 11-11.
The architecture on the left represents the traditional layered architecture, separated
by technical capabilities such as persistence, business rules, and so on. On the right,
the same solution appears, but separated by domain concerns such as Catalog Check
out and Update Inventory rather than technical capabilities.
Choreography Communication Style
|
309
Figure 11-11. Technical versus domain partitioning in architecture
Both topologies are logical ways to organize a codebase. However, consider where
domain concepts such as Catalog Checkout reside within each architecture, illustra‐
ted in Figure 11-12.
Figure 11-12. Catalog Checkout is smeared across implementation layers in a technically
partitioned architecture
310
|
Chapter 11: Managing Distributed Workflows
Catalog Checkout is “smeared” across the layers of the technical architecture,
whereas it appears only in the matching domain component and database in the
domain partitioned example. Of course, aligning a domain with domain partitioned
architecture isn’t a revelation—one of the insights of domain-driven design was the
primacy of the domain workflows. No matter what, if an architect wants to model a
workflow, they must make those moving parts work together. If the architect has
organized their architecture the same as the domains, the implementation of the
workflow should have similar complexity. However, if the architect has imposed addi‐
tional layers (as in technical partitioning, shown in Figure 11-12), it increases the
overall implementation complexity because now the architect must design for the
semantic complexity along with the additional implementation complexity.
Sometimes the extra complexity is warranted. For example, many layered architec‐
tures came from a desire by architects to gain cost savings by consolidating on archi‐
tecture patterns, such as database connection pooling. In that case, an architect
considered the trade-offs of the cost saving associated with technically partitioning
database connectivity versus the imposed complexity and cost won in many cases.
The major lesson of the last decade of architecture design is to model the semantics of
the workflow as closely as possible with the implementation.
An architect can never reduce semantic
implementation, but they can make it worse.
coupling
via
Thus, we can establish a relationship between the semantic coupling and the need for
coordination—the more steps required by the workflow, the more potential error and
other optional paths appear.
Workflow State Management
Most workflows include transient state about the status of the workflow: what ele‐
ments have executed, which ones are left, ordering, error conditions, retries, and so
on. For orchestrated solutions, the obvious workflow state owner is the orchestrator
(although some architectural solutions create stateless orchestrators for higher scale).
However, for choreography, no obvious owner for workflow state exists. Many com‐
mon options exist to manage state in choreography; here are three common ones.
First, the Front Controller pattern places the responsibility for state on the first called
service in the chain of responsibility, which in this case is Order Placement Service. If
that service contains information about both orders and the state of the workflow,
some of the domain services must have a communication link to query and update
the order state, as illustrated in Figure 11-13.
Choreography Communication Style
|
311
Figure 11-13. In choreography, a Front Controller is a domain service that owns
workflow state in addition to domain behavior
In this scenario, some services must communicate back to the Order Placement Ser‐
vice to update the state of the order, as it is the state owner. While this simplifies the
workflow, it increases communication overhead and makes the Order Placement Ser‐
vice more complex than one that handled only domain behavior. While the Front
Controller pattern has some advantageous characteristics, it also has trade-offs, as
shown in Table 11-2.
Trade-Offs
Table 11-2. Trade-offs for the Front Controller pattern
Advantage
Disadvantage
Creates a pseudo-orchestrator within
choreography
Makes querying the state of an order
trivial
Adds additional workflow state to a domain service
Increases communication overhead
Detrimental to performance and scale as it increases
integration communication chatter
312
|
Chapter 11: Managing Distributed Workflows
A second way for an architect to manage the transactional state is to keep no transi‐
ent workflow state at all, relying on querying the individual services to build a realtime snapshot. This is known as stateless choreography. While this simplifies the state
of the first service, it greatly increases network overhead in terms of chatter between
services to build a stateful snapshot. For example, consider a workflow like the simple
choreography happy path in Figure 11-7 with no extra state. If a customer wants to
know the state of their order, the architect must build a workflow that queries the
state of each domain service to determine the most up-to-date order status. While
this makes for a highly flexible solution, rebuilding state can be complex and costly in
terms of operational architecture characteristics like scalability and performance.
Stateless choreography trades high performance for workflow control, as illustrated in
Table 11-3.
Trade-Offs
Table 11-3. Stateless choreography trade-offs
Advantage
Disadvantage
Offers high performance and scale
Extremely decoupled
Workflow state must be built on the fly
Complexity rises swiftly with complex workflows
A third solution utilizes stamp coupling (described in more detail in “Stamp Coupling
for Workflow Management” on page 378), storing extra workflow state in the mes‐
sage contract sent between services. Each domain service updates its part of the over‐
all state and passes that to the next in the chain of responsibility. Thus, any consumer
of that contract can check on the status of the workflow without querying each
service.
This is a partial solution, as it still does not provide a single place for users to query
the state of the ongoing workflow. However, it does provide a way to pass the state
between services as part of the workflow, providing each service with additional
potentially useful context. As in all features of software architecture, stamp coupling
has good and bad characteristics, shown in Table 11-4.
Choreography Communication Style
|
313
Trade-Offs
Table 11-4. Stamp coupling trade-offs
Advantage
Disadvantage
Allows domain services to pass workflow state
without additional queries to a state owner
Eliminates need for a front controller
Contracts must be larger to accommodate
workflow state
Doesn’t provide just-in-time status queries
In Chapter 13, we discuss how contracts can reduce or increase workflow coupling in
choreographed solutions.
Advantages of the choreography communication style include the following:
Responsiveness
This communication style has fewer single choke points, thus offering more
opportunities for parallelism.
Scalability
Similar to responsiveness, lack of coordination points like orchestrators allows
more independent scaling.
Fault tolerance
The lack of a single orchestrator allows an architect to enhance fault tolerance
with the use of multiple instances.
Service decoupling
No orchestrator means less coupling.
Disadvantages of the choreography communication style include the following:
Distributed workflow
No workflow owner makes error management and other boundary conditions
more difficult.
State management
No centralized state holder hinders ongoing state management.
Error handling
Error handling becomes more difficult without an orchestrator because the
domain services must have more workflow knowledge.
Recoverability
Similarly, recoverability becomes more difficult without an orchestrator to
attempt retries and other remediation efforts.
314
|
Chapter 11: Managing Distributed Workflows
Like “Orchestration Communication Style” on page 301, choreography has a number
of good and bad trade-offs, often opposites, summarized in Table 11-5.
Trade-Offs
Table 11-5. Trade-offs for the choreography communication style
Advantage
Disadvantage
Responsiveness
Scalability
Fault tolerance
Service decoupling
Distributed workflow
State management
Error handling
Recoverability
Trade-Offs Between Orchestration and Choreography
As with all things in software architecture, neither orchestration nor choreography
represent the perfect solution for all possibilities. A number of key trade-offs, includ‐
ing some delineated here, will lead an architect toward one of these two solutions.
State Owner and Coupling
As illustrated in Figure 11-13, state ownership typically resides somewhere, either in a
formal mediator acting as an orchestrator, or a front controller in a choreographed
solution. In the choreographed solution, removing the mediator forces higher levels
of communication between services. This might be a perfectly suitable trade-off. For
example, if an architect has a workflow that needs higher scale and typically has few
error conditions, it might be worth trading the higher scale of choreography with the
complexity of error handling.
However, as workflow complexity goes up, the need for an orchestrator rises propor‐
tionally, as illustrated in Figure 11-14.
In addition, the more semantic complexity contained in a workflow, the more utili‐
tarian an orchestrator is. Remember, implementation coupling can’t make semantic
coupling better, only worse.
Trade-Offs Between Orchestration and Choreography
|
315
Ultimately, the sweet spot for choreography lies with workflows that need responsive‐
ness and scalability, and either don’t have complex error scenarios or they are infre‐
quent. This communication style allows for high throughput; it is used by the
dynamic coupling patterns “Phone Tag Saga(sac) Pattern” on page 330, “Time Travel
Saga(sec) Pattern” on page 336, and “Anthology Saga(aec) Pattern” on page 349. However,
it can also lead to extremely difficult implementations when other forces are mixed
in, leading to the “Horror Story(aac) Pattern” on page 343.
Figure 11-14. As the complexity of the workflow rises, orchestration becomes more useful
On the other hand, orchestration is best suited for complex workflows that include
boundary and error conditions. While this style doesn’t provide as much scale as
choreography, it greatly reduces complexity in most cases. This communication style
appears in “Epic Saga(sao) Pattern” on page 325, “Fairy Tale Saga(seo) Pattern” on page
333, “Fantasy Fiction Saga(aao) Pattern” on page 340, and “Parallel Saga(aeo) Pattern” on
page 346.
Coordination is one of the primary forces that create complexity for architects when
determining how to best communicate between microservices. Next, we investigate
how this force intersects with another primary force, consistency.
316
|
Chapter 11: Managing Distributed Workflows
Sysops Squad Saga: Managing Workflows
Thursday, March 15, 11:00
Addison and Austen arrived at Logan’s office right on time, armed with a presentation
and ritual coffee urn from the kitchen.
“Are you ready for us?” asked Addison.
“Sure,” said Logan. “Good timing—just got off a conference call. Are y’all ready to talk
about workflow options for the primary ticket flow?”
“Yes!” said Austen. “I think we should use choreography, but Addison thinks orchestration, and we can’t decide.”
“Give me an overview of the workflow we’re looking at.”
“It’s the primary ticket workflow,” said Addison. “It involves four services; here are the steps.”
Customer-facing operations
1. Customer submits a trouble ticket through the Ticket Management Service and receives a
ticket number.
Background operations
1. The Ticket Assignment Service finds the right Sysops expert for the trouble ticket.
2. The Ticket Assignment Service routes the trouble ticket to the systems expert’s mobile device.
3. The customer is notified via the Notification Service that the Sysops expert is on their way to fix
the problem.
4. The expert fixes the problem and marks the ticket as complete, which is sent to the Ticket Management Service.
5. The Ticket Management Service communicates with the Survey Service to tell the customer to
fill out the survey.
“Have you modeled both solutions?” asked Logan.
“Yes. The drawing for choreography is in Figure 11-15.”
Sysops Squad Saga: Managing Workflows
|
317
Figure 11-15. Primary ticket flow modeled as choreography
“…and the model for orchestration is in Figure 11-16.”
Figure 11-16. Primary ticket workflow modeled as orchestration
318
|
Chapter 11: Managing Distributed Workflows
Logan pondered the diagrams for a moment, then pronounced, “Well, there doesn’t seem to be an
obvious winner here. You know what that means.”
Austen piped up, “Trade-offs!”
“Of course,” laughed Logan. " Let’s think about the likely scenarios and see how each solution reacts
to them. What are the primary issues you are concerned with?”
“The first is lost or misrouted tickets. The business has been complaining about it, and it has become
a priority,” said Addison.
“OK, which handles that problem better—orchestration or choreography?”
“Easier control of the workflow sounds like the orchestrator version is better—we can handle all the
workflow issues there,” volunteered Austen.
“OK, let’s build a table of issues and preferred solutions in Table 11-6.”
Trade-Offs
Table 11-6. Trade-off between orchestration and choreography for ticket
workflow
Orchestration
Choreography
Workflow control
“What’s the next issue we should model?” Addison asked.
“We need to know the status of a trouble ticket at any given moment—the business has requested
this feature, and it makes it easier to track several metrics. That implies we need an orchestrator so
that we can query the state of the workflow.”
“But you don’t have to have an orchestrator for that—we can query any given service to see if it has
handled a particular part of the workflow, or use stamp coupling,” said Addison.
“That’s right—this isn’t a zero-sum game,” said Logan. “It’s possible that both or neither work just as
well. We’ll give both solutions credit in our updated table in Table 11-7.”
Sysops Squad Saga: Managing Workflows
|
319
Trade-Offs
Table 11-7. Updated trade-offs between orchestration and choreography for
ticket workflow
Orchestration
Choreography
Workflow control
State query
State query
“OK, what else?”
“Just one more that I can think of,” Addison said. “Tickets can get canceled by the customer, and tickets can get reassigned because of expert availability, lost connections to the expert’s mobile device,
or expert delays at a customer site. Therefore, proper error handling is important. That means
orchestration?”
“Yes, generally. Complex workflows must go somewhere, either in an orchestrator or scattered
through services. It’s nice to have a single place to consolidate error handling. And choreography
definitely does not score well here, so we’ll update our table in Table 11-8.”
Trade-Offs
Table 11-8. Final trade-offs between orchestration and choreography for ticket
workflow
Orchestration
Choreography
Workflow control
State query
State query
Error handling
“That looks pretty good. Any more?”
“Nothing that’s not obvious,” said Addison. “We’ll write this up in an ADR; in case we think of any
other issues, we can add them there.”
320
|
Chapter 11: Managing Distributed Workflows
ADR: Use Orchestration for Primary Ticket Workflow
Context
For the primary ticket workflow, the architecture must support easy tracking of lost or mistracked messages, excellent error handling, and the ability to track ticket status. Either an
orchestration solution illustrated in Figure 11-16 or a choreography solution illustrated in
Figure 11-15 will work.
Decision
We will use orchestration for the primary ticketing workflow.
We modeled orchestration and choreography and arrived at the trade-offs in Table 11-8.
Consequences
Ticketing workflow might have scalability issues around a single orchestrator, which
should be reconsidered if current scalability requirements change.
Sysops Squad Saga: Managing Workflows
|
321
CHAPTER 12
Transactional Sagas
Thursday, March 31, 16:55
Austen showed up at Logan’s office late on a windy Thursday afternoon. “Addison just
sent me over here to ask you about some horror story?”
Logan stopped and looked up. “Is that a description of whatever crazy extreme sport
you’re doing this weekend? What is it this time?”
“It’s late spring, so a bunch of us are going ice skating on the thawing lake. We’re wearing
body suits, so it’s really a combination of skating and swimming. But that’s not what
Addison meant at all. When I showed Addison my design for the Ticketing workflow, I was immediately instructed to come to you and tell you I’ve created a horror story.”
Logan laughed. “Oh, I see what’s going on—you stumbled into the Horror Story saga communication pattern. You designed a workflow with asynchronous communication, atomic transactionality,
and choreography, right?”
“How did you know?”
“That’s the Horror Story saga pattern, or really, anti-pattern. There are eight generic saga patterns we
start from, so it’s good to know what they are, because each has a different balance of trade-offs.”
The concept of a saga in architecture predates microservices, originally concerned
with limiting the scope of database locks in early distributed architectures—the paper
largely assumed to have coined the concept is from the Proceedings of the 1987 ACM
conference. In his book Microservices Patterns (Manning Publications) and also out‐
lined in the “Saga Pattern” section of his website, Chris Richardson describes the saga
pattern for microservices as a sequence of local transactions where each update pub‐
lishes an event, thus triggering the next update in the sequence. If any of those
323
updates fail, the saga issues a series of compensating updates to undo the prior
changes made during the saga.
However, recall from Chapter 2 that this is only one of eight possible types of sagas.
In this section, we dive much deeper and look at the inner workings of transactional
sagas and how to manage them, particularly when errors occur. After all, since dis‐
tributed transactions lack atomicity (see “Distributed Transactions” on page 263),
what makes them interesting is when problems occur.
Transactional Saga Patterns
In Chapter 2, we introduced a matrix that juxtaposed each of the intersecting dimen‐
sions when architects must choose how to implement a transactional saga, repro‐
duced in Table 12-1.
Table 12-1. The matrix of dimensional intersections for distributed architectures
Pattern name
Communication
Consistency
Coordination
Epic Saga(sao)
Phone Tag Saga(sac)
Fairy Tale Saga(seo)
Time Travel Saga(sec)
Fantasy Fiction Saga(aao)
Horror Story(aac)
Parallel Saga(aeo)
Anthology Saga(aec)
Synchronous
Synchronous
Synchronous
Synchronous
Asynchronous
Asynchronous
Asynchronous
Asynchronous
Atomic
Atomic
Eventual
Eventual
Atomic
Atomic
Eventual
Eventual
Orchestrated
Choreographed
Orchestrated
Choreographed
Orchestrated
Choreographed
Orchestrated
Choreographed
We provide whimsical names for each combination, all derived from types of sagas.
However, the pattern names exist to help differentiate the possibilities, and we don’t
want to provide a memorization test to associate a pattern name to a set of character‐
istics, so we have added a superscript to each saga type indicating the values of the
three dimensions listed in alphabetical order (as in Table 12-1). For example, the Epic
Saga(sao) pattern indicates the values of synchronous, atomic, and orchestrated for com‐
munication, consistency, and coordination. The superscripts help you associate
names to character sets more easily.
While architects will utilize some of the patterns more than others, they all have legit‐
imate uses and differing sets of trade-offs.
324
| Chapter 12: Transactional Sagas
We illustrate each possible communication combination with both a threedimensional representation of the intersection of the three forces in space along with
an example workflow using generic distributed services, which we refer to as isomor‐
phic diagrams. These diagrams show interactions between services in the most
generic way, toward our goal of showing architect concepts in the simplest form. In
each of these diagrams, we use the set of generic symbols shown in Figure 12-1.
Figure 12-1. Legend for ISO architecture interaction diagrams
For each of the architecture patterns, we do not show every possible interaction,
which would become repetitive. Instead, we identify and illustrate the differentiating
features of the pattern—what makes its behavior unique among the patterns.
Epic Saga(sao) Pattern
This type of communication is the “traditional” saga pattern as many architects
understand it, also called an Orchestrated Saga because of its coordination type. Its
dimensional relationships appear in Figure 12-2.
This pattern utilizes synchronous communication, atomic consistency, and orchestra‐
ted coordination. The architect’s goal when choosing this pattern mimics the behavior
of monolithic systems—in fact, if a monolithic system were added to this diagram in
Figure 12-2, it would be the origin (0, 0, 0), lacking distribution entirely. Thus, this
communication style is most familiar with architects and developers of traditional
transactional systems.
Transactional Saga Patterns
|
325
Figure 12-2. The Epic Saga(sao) pattern’s dynamic coupling (communication, consistency,
coordination) relationships
The isomorphic representation of the Epic Saga(sao) pattern appears in Figure 12-3.
Figure 12-3. The isomorphic communication illustration of the Epic Saga(sao) pattern
326
|
Chapter 12: Transactional Sagas
Here, an orchestrator service orchestrates a workflow that includes updates for three
services, expected to occur transactionally—either all three calls succeed or none do.
If one of the calls fails, they all fail and return to the previous state. An architect can
solve this coordination problem in a variety of ways, all complex in distributed archi‐
tectures. However, such transactions limit the choice of databases and have legendary
failure modes.
Many nascent or naive architects trust that, because a pattern exists for a problem, it
represents a clean solution. However, the pattern is recognition of only commonality,
not solvability. Distributed transactions provide an excellent example of this phe‐
nomenon—architects accustomed to modeling transactions in nondistributed sys‐
tems sometimes believe that moving that capability to the distributed world is an
incremental change. However, transactions in distributed architectures present a
number of challenges, which become proportionally worse depending on the com‐
plexity of the semantic coupling of the problem.
Consider a common implementation of the Epic Saga(sao) pattern, utilizing compen‐
sating transactions. A compensating update is one that reverses a data write action
performed by another service (such as reversing an update, reinserting a previously
deleted row, or deleting a previously inserted row) during the course of the dis‐
tributed transaction scope. While compensating updates attempt to reverse changes
in order to bring distributed data sources back to their original state prior to the start
of the distributed transaction, they are riddled with complex issues, challenges, and
trade-offs.
A compensating transaction pattern assigns a service to monitor the transactional
completeness of a request, as shown in Figure 12-4.
Figure 12-4. A successful orchestrated transactional Epic Saga using a compensating
transaction
Transactional Saga Patterns
|
327
However, as with many things in architecture, the error conditions cause the difficul‐
ties. In a compensating transaction framework, the mediator monitors the success of
calls, and issues compensating calls to other services if one or more of the requests
fail, as shown in Figure 12-5.
Figure 12-5. When an error occurs, a mediator must send compensating requests to
other services
A mediator both accepts requests and mediates the workflow, and synchronous calls
to the first two services succeed. However, when trying to make the call to the last
service, it fails (from a possibly a wide variety of both domain and operational rea‐
sons). Because the goal of the Epic Saga(sao) is atomic consistency, the mediator must
utilize compensating transactions and request that the other two services undo the
operation from before, returning the overall state to what it was before the transac‐
tion started.
This pattern is widely used: it models familiar behavior, and it has a well-established
pattern name. Many architects default to the Epic Saga(sao) pattern because it feels
familiar to monolithic architectures, combined with a request (sometimes demand)
from stakeholders that state changes must synchronize, regardless of technical con‐
straints. However, many of the other dynamic quantum coupling patterns may offer a
better set of trade-offs.
The clear advantage of the Epic Saga(sao) is the transactional coordination that mimics
monolithic systems, coupled with the clear workflow owner represented via an
orchestrator. However, the disadvantages are varied. First, orchestration plus transac‐
tionality may have an impact on operational architecture characteristics such as per‐
formance, scale, elasticity, and so on—the orchestrator must make sure that all
participants in the transaction have succeeded or failed, creating timing bottlenecks.
Second, the various patterns used to implement distributed transactionality (such as
328
|
Chapter 12: Transactional Sagas
compensating transactions) succumb to a wide variety of failure modes and bound‐
ary conditions, along with adding inherent complexity via undo operations. Dis‐
tributed transactions present a host of difficulties and thus are best avoided if
possible.
The Epic Saga(sao) pattern features the following characteristics:
Coupling level
This pattern exhibits extremely high levels of coupling across all possible dimen‐
sions: synchronous communication, atomic consistency, and orchestrated coor‐
dination—it is in fact the most highly coupled pattern in the list. This isn’t
surprising, as it models the behavior of highly coupled monolithic system com‐
munication, but creates a number of issues in distributed architectures.
Complexity level
Error conditions and other intensive coordination added to the requirement of
atomicity add complexity to this architecture. The synchronous calls this archi‐
tecture uses mitigate some of the complexity, as architects don’t have to worry
about race conditions and deadlocks during calls.
Responsiveness/availability
Orchestration creates a bottleneck, especially when it must also coordinate trans‐
actional atomicity, which reduces responsiveness. This pattern uses synchronous
calls, further impacting performance and responsiveness. If any of the services
are not available or an unrecoverable error occurs, this pattern will fail.
Scale/elasticity
Similar to responsiveness, the bottleneck and coordination required to implement
this pattern make scale and other operational concerns difficult.
While the Epic Saga(sao) is popular because of familiarity, it creates a number of chal‐
lenges, both from a design and operational characteristics standpoint, as shown in
Table 12-2.
Table 12-2. Ratings for the Epic Saga(sao)
Epic Saga(sao) pattern
Ratings
Communication
Consistency
Coordination
Coupling
Complexity
Responsiveness/availability
Scale/elasticity
Synchronous
Atomic
Orchestrated
Very high
Low
Low
Very low
Transactional Saga Patterns
|
329
Fortunately, architects need not default to patterns that, while seemingly familiar, cre‐
ate accidental complexity—a variety of other patterns exist with differing sets of
trade-offs. Refer to the “Sysops Squad Saga: Atomic Transactions and Compensating
Updates” on page 358 for a concrete example of the Epic Saga(sao) and some of the
complex challenges it presents (and how to address those challenges).
Phone Tag Saga(sac) Pattern
The Phone Tag Saga(sac) pattern changes one of the dimensions of the Epic Saga(sao),
changing coordination from orchestrated to choreographed; this change is illustrated in
Figure 12-6.
Figure 12-6. The Phone Tag pattern utilizes loosely coupled communication
The pattern name is Phone Tag because it resembles a well-known children’s game
known as Telephone in North America: children form a circle, and one person whis‐
pers a secret to the next person, who passes it along to the next, until the final version
is spoken by the last person. In Figure 12-6, choreography is favored over orchestra‐
tion, creating the corresponding change in the structural communication shown in
Figure 12-7.
330
| Chapter 12: Transactional Sagas
Figure 12-7. Because of a lack of orchestration, each participant must coordinate status
The Phone Tag Saga(sac) pattern features atomicity but also choreography, meaning
that the architect designates no formal orchestrator. Yet atomicity requires some
degree of coordination. In Figure 12-7, the initially called service becomes the coordi‐
nation point (sometimes called the front controller). Once it has finished its work, it
passes a request on to the next service in the workflow, which continues until the
workflow succeeds. However, if an error condition occurs, each service must have
built-in logic to send compensating requests back along the chain.
Because the architectural goal is transactional atomicity, logic to coordinate that
atomicity must reside somewhere. Thus, domain services must contain more logic
about the workflow context they participate within, including error handling and
routing. For complex workflows, the front controller in this pattern will become as
complex as most mediators, reducing the appeal and applicability of this pattern.
Thus, this pattern is commonly used for simple workflows that need higher scale, but
with a potential performance impact.
How does choreography versus orchestration improve operational architecture char‐
acteristics like scale? Using choreography even with synchronous communication
cuts down on bottlenecks—in nonerror conditions, the last service in the workflow
can return the result, allowing for higher throughput and fewer choke points. Perfor‐
mance for happy path workflows can be faster than in an Epic Saga(sao) because of
lack of coordination. However, error conditions will be much slower without a medi‐
ator—each service must unwind the call chain, which also increases coupling between
services.
Transactional Saga Patterns
|
331
Generally, the Phone Tag Saga(sac) offers slightly better scale than the Epic Saga(sao)
because of the lack of a mediator, which can sometimes become a limiting bottleneck.
However, this pattern also features lower performance for error conditions and other
workflow complexities—without a mediator, the workflow must be resolved via com‐
munication between services, which impacts performance.
A nice feature of nonorchestrated architectures is the lack of a coupling singularity, a
single place the workflow couples to. Even though this pattern utilizes synchronous
requests, fewer wait conditions for happy path workflows exist, allowing for higher
scale. In general, reducing coupling increases scale.
With the improved scalability brought about because of a lack of orchestration comes
the increased complexity of the domain services to manage the workflow concerns in
addition to their nominal responsibility. For complex workflows, increased complex‐
ity and interservice communication may drive architects back toward orchestration
and its trade-offs.
The Phone Tag Saga(sac) has a fairly rare combination of features—generally, if an
architect chooses choreography, they also choose asynchronicity. However, in some
cases where an architect might choose this combination instead: synchronous calls
ensure that each domain service completes its part of the workflow before invoking
the next, eliminating race conditions. If error conditions are easy to resolve, or
domain services can utilize idempotence and retries, then architects can build higher
parallel scale using this pattern compared to an Epic Saga(sao).
The Phone Tag Saga(sac) pattern has the following characteristics:
Coupling level
This pattern relaxes one of the coupling dimensions of the Epic Saga(sao) pattern,
utilizing a choreographed rather than orchestrated workflow. Thus, this pattern is
slightly less coupled, but with the same transactional requirement, meaning that
the complexity of the workflow must be distributed between the domain services.
Complexity level
This pattern is significantly more complex than the Epic Saga(sao); complexity in
this pattern rises linearly proportionally to the semantic complexity of the work‐
flow: the more complex the workflow, the more logic must appear in each service
to compensate for lack of orchestrator. Alternatively, an architect might add
workflow information to the messages themselves as a form of stamp coupling
(see “Stamp Coupling for Workflow Management” on page 378) to maintain state
but adding to the overhead context required by each service.
332
|
Chapter 12: Transactional Sagas
Responsiveness/availability
Less orchestration generally leads to better responsiveness, but error conditions
in this pattern become more difficult to model without an orchestrator, requiring
more coordination via callbacks and other time-consuming activities.
Scale/elasticity
Lack of orchestration translates to fewer bottlenecks, generally increasing scala‐
bility, but only slightly. This pattern still utilizes tight coupling around two of the
three dimensions, so scalability isn’t a highlight, especially if error conditions are
common.
The ratings for the Phone Tag Saga(sac) appear in Table 12-3.
Table 12-3. Ratings for the Phone Tag Saga(sac)
Phone Tag Saga(sac)
Ratings
Communication
Consistency
Coordination
Coupling
Complexity
Responsiveness/availability
Scale/elasticity
Synchronous
Atomic
Choreographed
High
High
Low
Low
The Phone Tag Saga(sac) pattern is better for simple workflows that don’t have many
common error conditions. While it offers a few better characteristics than the Epic
Saga(sao), the complexity introduced by lack of an orchestrator offsets many of the
advantages.
Fairy Tale Saga(seo) Pattern
Typical fairy tales provide happy stories with easy-to-follow plots, thus the name
Fairy Tale Saga(seo), which utilizes synchronous communication, eventual consistency,
and orchestration, as shown in Figure 12-8.
Transactional Saga Patterns
|
333
Figure 12-8. The Fairy Tale Saga(seo) illustrates eventual consistency
This communication pattern relaxes the difficult atomic requirement, providing
many more options for architects to design systems. For example, if a service is down
temporarily, eventual consistency allows for caching a change until the service
restores. The communication structure for the Fairy Tale Saga(seo) is illustrated in Fig‐
ure 12-9.
In this pattern, an orchestrator exists to coordinate request, response, and error han‐
dling. However, the orchestrator isn’t responsible for managing transactions, which
each domain service retains responsibility for (for examples of common workflows,
see Chapter 11). Thus the orchestrator can manage compensating calls, but without
the requirement of occurring within an active transaction.
334
|
Chapter 12: Transactional Sagas
Figure 12-9. Isomorphic illustration of a Fairy Tale interaction
This is a much more attractive pattern and appears commonly in many microservices
architectures. Having a mediator makes managing workflows easier, synchronous
communication is the easier of the two choices, and eventual consistency removes the
most difficult coordination challenge, especially for error handling.
The biggest appealing advantage of the Fairy Tale Saga(seo) is the lack of holistic trans‐
actions. Each domain service manages its own transactional behavior, relying on
eventual consistency for the overall workflow.
Compared to many other patterns, this pattern generally exhibits a good balance of
trade-offs:
Coupling level
The Fairy Tale Saga(seo) features high coupling, with two of the three coupling
drivers maximized in this pattern (synchronous communication and orchestra‐
ted coordination). However, the worse driver of coupling complexity—transac‐
tionality—disappears in this pattern in favor of eventual consistency. The
orchestrator must still manage complex workflows, but without the stricture of
doing so within a transaction.
Complexity level
Complexity for the Fairy Tale Saga(seo) is quite low; it includes the most conve‐
nient options (orchestrated, synchronicity) with the loosest restriction (eventual
consistency). Thus the name Fairy Tale Saga(seo)—a simple story with a happy
ending.
Transactional Saga Patterns
|
335
Responsiveness/availability
Responsiveness is typically better in communication styles of this type because,
even though the calls are synchronous, the mediator needs to contain less timesensitive state about ongoing transactions, allowing for better load balancing.
However, true distinctions in performance come with asynchronicity, illustrated
in future patterns.
Scale/elasticity
Lack of coupling generally leads to higher scale; removing transactional coupling
allows each service to scale more independently.
The ratings for the Fairy Tale Saga(seo) appear in Table 12-4.
Table 12-4. Ratings for the Fairy Tale Saga(seo)
Fairy Tale Saga(seo)
Ratings
Communication
Consistency
Coordination
Coupling
Complexity
Responsiveness/availability
Scale/elasticity
Synchronous
Eventual
Orchestrated
High
Very low
Medium
High
If an architect can take advantage of eventual consistency, this pattern is quite attrac‐
tive, combining the easy moving parts with the fewest scary restrictions, making it a
popular choice among architects.
Time Travel Saga(sec) Pattern
The Time Travel Saga(sec) pattern features synchronous communication, and eventual
consistency, but choreographed workflow. In other words, this pattern avoids a cen‐
tral mediator, placing the workflow responsibilities entirely on the participating
domain services, as illustrated in Figure 12-10.
336
|
Chapter 12: Transactional Sagas
Figure 12-10. The Time Travel Saga(sec) pattern uses two of three decoupling techniques
The structural topology illustrates the lack of orchestration, shown in Figure 12-11.
Figure 12-11. Complex workflows become difficult to manage without orchestration
Transactional Saga Patterns
|
337
In this workflow, each service accepts a request, performs an action, and then for‐
wards the request on to another service. This architecture can implement the Chain
of Responsibility design pattern or the Pipes and Filters architecture style—any work‐
flow with a one-way series of steps. Each service in this pattern “owns” its own trans‐
actionality, so architects must design workflow error conditions into the domain
design. In general, a proportional complexity relationship exists between workflow
complexity and choreographed solutions because of a lack of built-in coordination
via a mediator—the more complex the workflow, the more difficult choreography
becomes. It is called Time Travel Saga(sec) because everything is decoupled from a
time standpoint: each service owns its own transactional context, making workflow
consistency temporally gradual—the state will become consistent over time based on
the design of the interaction.
The lack of transactions in the Time Travel Saga(sec) pattern makes workflows easier
to model; however, the lack of an orchestrator means that each domain service must
include most workflow state and information. As in all choreographed solutions, a
direct correlation exists between workflow complexity and the utility of an orchestra‐
tor; thus, this pattern is best suited for simple workflows.
For solutions that benefit from high throughput, this pattern works extremely well for
“fire and forget” style workflows, such as electronic data ingestion, bulk transactions,
and so on. However, because no orchestrator exists, domain services must deal with
error conditions and coordination.
Lack of coupling increases scalability with this pattern; only adding asynchronicity
would make it more scalable (as in the Anthology Saga(aec) pattern). However,
because this pattern lacks holistic transactional coordination, architects must take
extra effort to synchronize data.
Here is the qualitative evaluation of the Time Travel Saga(sec) pattern:
Coupling level
The coupling level falls in the medium range with the Time Travel Saga(sec), with
the decreased coupling brought on by the absence of an orchestrator balanced by
the still remaining coupling of synchronous communication. As with all eventual
consistency patterns, the absence of transactional coupling eases many data
concerns.
Complexity level
The loss of transactionality provides a decrease in complexity for this pattern.
This pattern is quasi-special-purpose, superbly suited to fast throughput, oneway communication architectures, and the coupling level matches that style of
architecture well.
338
|
Chapter 12: Transactional Sagas
Responsiveness/availability
Responsiveness scores a medium with this architectural pattern: it is quite high for
built-to-purpose systems, as described previously, and quite low for complex
error handling. Because no orchestrator exists in this pattern, each domain ser‐
vice must handle the scenario to restore eventual consistency in the case of an
error condition, which will cause a lot of overhead with synchronous calls,
impacting responsiveness and performance.
Scale/elasticity
This architecture pattern offers extremely good scale and elasticity; it could only
be made better with asynchronicity (see the Anthology Saga(aec) pattern).
The ratings for the Time Travel Saga(sec) pattern appear in Table 12-5.
Table 12-5. Ratings for the Time Travel Saga(sec)
Time Travel Saga(sec)
Ratings
Communication
Consistency
Coordination
Coupling
Complexity
Responsiveness/availability
Scale/elasticity
Synchronous
Eventual
Choreographed
Medium
Low
Medium
High
The Time Travel Saga(sec) pattern provides an on-ramp to the more complex but ulti‐
mately scalable Anthology Saga(aec) pattern. Architects and developers find dealing
with synchronous communication easier to reason about, implement, and debug; if
this pattern provides adequate scalability, teams don’t have to embrace the more com‐
plex but more scalable alternatives.
Transactional Saga Patterns
|
339
Fantasy Fiction Saga(aao) Pattern
The Fantasy Fiction Saga(aao) uses atomic consistency, asynchronous communication,
and orchestrated coordination, as shown in Figure 12-12.
Figure 12-12. Asynchronous communication makes transactionality difficult in this pat‐
tern
The structure representation shown in Figure 12-13 starts to show some of the diffi‐
culties with this pattern.
Just because a combination of architectural forces exists doesn’t mean it forms an
attractive pattern, yet this relatively implausible combination has uses. This pattern
resembles the Epic Saga(sao) in all aspects except for communication—this pattern uses
asynchronous rather than synchronous communication. Traditionally, one way that
architects increase the responsiveness of distributed systems is by using asynchronic‐
ity, allowing operations to occur in parallel rather than serially. This may seem like a
good way to increase the perceived performance over an Epic Saga(sao).
340
|
Chapter 12: Transactional Sagas
Figure 12-13. The Fantasy Fiction Saga(aao) pattern is far-fetched because transaction
coordination for asynchronous communication presents difficulties
However, asynchronicity isn’t a simple change—it adds many layers of complexity to
architecture, especially around coordination, requiring much more complexity in the
mediator. For example, suppose a transactional workflow Alpha begins. Because
everything is asynchronous, while Alpha is pending, transactional workflow Beta
begins. Now, the mediator must keep track of the state of all ongoing transactions in
pending state.
It gets worse. Suppose that workflow Gamma begins, but the first call to the domain
service depends on the still pending outcome of Alpha—how can an architect model
this behavior? While possible, the complexity grows and grows.
Adding asynchronicity to orchestrated workflows adds asynchronous transactional
state to the equation, removing serial assumptions about ordering and adding the
possibilities of deadlocks, race conditions, and a host of other parallel system
challenges.
Transactional Saga Patterns
|
341
This pattern offers the following challenges:
Coupling level
The coupling level is extremely high in this pattern, using an orchestrator and
atomicity but with asynchronous communication, which makes coordination
more difficult because architects and developers must deal with race conditions
and other out-of-order problems imposed by asynchronous communication.
Complexity level
Because the coupling is so difficult, the complexity rises in this pattern as well.
There’s not only design complexity, requiring architects to develop overly com‐
plex workflows, but also debugging and operational complexity of dealing with
asynchronous workflows at scale.
Responsiveness/availability
Because this pattern attempts transactional coordination across calls, responsive‐
ness will be impacted overall and be extremely bad if one or more of the services
isn’t available.
Scale/elasticity
High scale is virtually impossible in transaction systems, even with asynchronic‐
ity. Scale is much better in the similar pattern Parallel Saga(aeo), which switches
atomic to eventual consistency.
The ratings for the Fantasy Fiction Saga(aao) pattern appear in Table 12-6.
Table 12-6. Ratings for the Fantasy Fiction Saga(aao)
Fantasy Fiction
Ratings
Communication
Consistency
Coordination
Coupling
Complexity
Responsiveness/availability
Scale/elasticity
Asynchronous
Atomic
Orchestrated
High
High
Low
Low
This pattern is unfortunately more popular than it should be, mostly from the misguided attempt to improve the performance of Epic Saga(sao) while maintaining trans‐
actionality; a better option is usually Parallel Saga(aeo).
342
|
Chapter 12: Transactional Sagas
Horror Story(aac) Pattern
One of the patterns must be the worst possible combination; it is the aptly named
Horror Story(aac) pattern, characterized by asynchronous communication, atomic con‐
sistency, and choreographed coordination, illustrated in Figure 12-14.
Figure 12-14. The most difficult combination: achieving transactionality while asynchro‐
nous and choreographed
Why is this combination so horrible? It combines the most stringent coupling around
consistency (atomic) with the two loosest coupling styles, asynchronous and choreog‐
raphy. The structural communication for this pattern appears in Figure 12-15.
Transactional Saga Patterns
|
343
Figure 12-15. This pattern requires a lot of interservice communication because of
required transactionality and the lack of a mediator
In this pattern, no mediator exists to manage transactional consistency across multi‐
ple services—while using asynchronous communication. Thus, each domain service
must track undo information about multiple pending transactions, potentially out of
order because of asynchronicity, and coordinate with each other during error condi‐
tions. For just one of many possible horrible examples, imagine that transaction
Alpha starts and, while pending, transaction Beta starts. One of the calls for the Alpha
transaction fails—now, the choreographed services have to reverse the order of firing,
undoing each (potentially out-of-order) element of the transaction along the way.
The multiplicity and complexity of error conditions makes this a daunting option.
Why might an architect choose this option? Asynchronicity is appealing as a perfor‐
mance boost, yet the architect may still try to maintain transactional integrity, which
has many myriad failure modes. Instead, an architect would be better off choosing the
Anthology Saga(aec) pattern, which removes holistic transactionality.
The qualitative evaluations for the Horror Story(aac) pattern are as follows:
Coupling level
Surprisingly, the coupling level for this pattern isn’t the worst (that “honor” goes
to the Epic Saga(sao) pattern). While this pattern does attempt the worst kind of
single coupling (transactionality), it relieves the other two, lacking both a media‐
tor and the coupling—increasing synchronous communication.
344
|
Chapter 12: Transactional Sagas
Complexity level
Just as the name implies, the complexity of this pattern is truly horrific, the worst
of any because it requires the most stringent requirement (transactionality) with
the most difficult combination of other factors to achieve that (asynchronicity
and choreography).
Scale/elasticity
This pattern does scale better than ones with a mediator, and asynchronicity also
adds the ability to perform more work in parallel.
Responsiveness/availability
Responsiveness is low for this pattern, similar to the other patterns that require
holistic transactions: coordination for the workflow requires a large amount of
interservice “chatter,” hurting performance and responsiveness.
The trade-offs for the Horror Story(aac) pattern appear in Table 12-7.
Table 12-7. Ratings for the Horror Story(aac)
Horror Story(aac)
Ratings
Communication
Consistency
Coordination
Coupling
Complexity
Responsiveness/availability
Scale/elasticity
Asynchronous
Atomic
Choreographed
Medium
Very high
Low
Medium
The aptly named Horror Story(aac) pattern is often the result of a well-meaning archi‐
tect starting with an Epic Saga(sao) pattern, noticing slow performance because of
complex workflows, and realizing that techniques to improve performance include
asynchronous communication and choreography. However, this thinking provides an
excellent example of not considering all the entangled dimensions of a problem
space. In isolation, asynchronous communication improves performance. However,
as architects, we cannot consider it in isolation when it is entangled with other archi‐
tecture dimensions, such as consistency and coordination.
Transactional Saga Patterns
|
345
Parallel Saga(aeo) Pattern
The Parallel Saga(aeo) pattern is named after the “traditional” Epic Saga(sao) pattern
with two key differences that ease restrictions and therefore make it an easier pattern
to implement: asynchronous communication and eventual consistency. The dimen‐
sional diagram of the Parallel Saga(aeo) pattern appears in Figure 12-16.
Figure 12-16. Parallel Saga(aeo) offers performance improvements over traditional sagas
The most difficult goals in the Epic Saga(sao) pattern revolve around transactions and
synchronous communication, both of which cause bottlenecks and performance deg‐
radation. As shown in Figure 12-16, the pattern loosens both restraints.
The isomorphic representation of Parallel Saga(aeo) appears in Figure 12-17.
This pattern uses a mediator, making it suitable for complex workflows. However, it
uses asynchronous communication, allowing for better responsiveness and parallel
execution. Consistency in the pattern lies with the domain services, which may
require some synchronization of shared data, either in the background or driven via
the mediator. As in other architectural problems that require coordination, a media‐
tor becomes quite useful.
346
|
Chapter 12: Transactional Sagas
Figure 12-17. Each service owns its own transactionality; the mediator coordinates
request and response
For example, if an error occurs during the execution of a workflow, the mediator can
send asynchronous messages to each involved domain service to compensate for the
failed change, which may entail retries, data synchronization, or a host of other
remediations.
Of course, the loosening of constraints implies that some benefits will be traded off,
which is the nature of software architecture. Lack of transactionality imposes more
burden on the mediator to resolve error and other workflow issues. Asynchronous
communication, while offering better responsiveness, makes resolving timing and
synchronization issues difficult—race conditions, deadlocks, queue reliability, and a
host of other distributed architecture headaches reside in this space.
The Parallel Saga(aeo) pattern exhibits the following qualitative scores:
Coupling level
This pattern has a low coupling level, isolating the coupling-intensifying force of
transactions to the scope of the individual domain services. It also utilizes asyn‐
chronous communication, further decoupling services from wait states, allowing
for more parallel processing but adding a time element to an architect’s coupling
analysis.
Complexity level
The complexity of the Parallel Saga(aeo) is also low, reflecting the lessening of cou‐
pling stated previously. This pattern is fairly easy for architects to understand,
and orchestration allows for simpler workflow and error-handling designs.
Transactional Saga Patterns
|
347
Scale/elasticity
Using asynchronous communication and smaller transaction boundaries allows
this architecture to scale quite nicely, and with good levels of isolation between
services. For example, in a microservices architecture, some public-facing serv‐
ices might need higher levels of scale and elasticity, where back office services
don’t need scale but higher levels of security. Isolating transactions at the domain
level frees the architecture to scale around domain concepts.
Responsiveness/availability
Because of lack of coordinated transactions and asynchronous communication,
the responsiveness of this architecture is high. In fact, because each of these serv‐
ices maintains its own transactional context, this architecture is well suited to
highly variable service performance footprints between services, allowing archi‐
tects to scale some services more than others because of demand.
The ratings associated with the Parallel Saga(aeo) pattern appear in Table 12-8.
Table 12-8. Ratings for the Parallel Saga(aeo)
Parallel Saga(aeo)
Ratings
Communication
Consistency
Coordination
Coupling
Complexity
Responsiveness/availability
Scale/elasticity
Asynchronous
Eventual
Orchestrated
Low
Low
High
High
Overall, the Parallel Saga(aeo) pattern offers an attractive set of trade-offs for many
scenarios, especially with complex workflows that need high scale.
348
|
Chapter 12: Transactional Sagas
Anthology Saga(aec) Pattern
The Anthology Saga(aec) pattern provides the exact opposite set of characteristics to
the traditional Epic Saga(sao) pattern: it utilizes asynchronous communication, even‐
tual consistency, and choreographed coordination, providing the least coupled exem‐
plar among all these patterns. The dimensional view of the Anthology Saga(aec)
pattern appears in Figure 12-18.
Figure 12-18. The Anthology Saga(aec) pattern offers the opposite extremes of the Epic
Saga, and is therefore the least coupled pattern
The anthology pattern uses message queues to send asynchronous messages to other
domain services without orchestration, as illustrated in Figure 12-19.
As you can see, each service maintains its own transactional integrity, and no orches‐
trator exists, forcing each domain service to include more context about the work‐
flows they participate in, including error handling and other coordination strategies.
Transactional Saga Patterns
|
349
Figure 12-19. Lack of orchestration, eventual consistency, and asynchronicity make this
pattern highly decoupled but a challenge for coordination
The lack of orchestration makes services more complex but allows for much higher
throughput, scalability, elasticity, and other beneficial operational architecture char‐
acteristics. No bottlenecks or coupling choke points exist in this architecture, allow‐
ing for high responsiveness and scalability.
However, this pattern doesn’t work particularly well for complex workflows, espe‐
cially around resolving data consistency errors. While it may not seem possible
without an orchestrator, stamp coupling (“Stamp Coupling for Workflow Manage‐
ment” on page 378) may be used to carry workflow state, as described in the similar
Phone Tag Saga(sac) pattern.
This pattern works best for simple, mostly linear workflows, where architects desire
high processing throughput. This pattern provides the most potential for both high
performance and scale, making it an attractive choice when those are key drivers for
the system. However, the degree of decoupling makes coordination difficult, prohibi‐
tively so for complex or critical workflows.
The short-story-inspired
characteristics:
Anthology
Saga(aec)
pattern
has
the
following
Coupling level
Coupling for this pattern is the lowest for any other combination of forces, creat‐
ing a highly decoupled architecture well suited for high scale and elasticity.
Complexity level
While the coupling is extremely low, complexity is correspondingly high, espe‐
cially for complex workflows where an orchestrator (lacking here) is convenient.
350
| Chapter 12: Transactional Sagas
Scale/elasticity
This pattern scores the highest in the scale and elasticity category, correlating
with the overall lack of coupling found in this pattern.
Responsiveness
Responsiveness is high in this architecture because of a lack of speed governors
(transactional consistency, synchronous communication) and use of responsive‐
ness accelerators (choreographed coordination).
The ratings table for the Anthology Saga(aec) pattern appear in Table 12-9.
Table 12-9. Ratings for the Anthology Saga(aec)
Anthology Saga(aec)
Ratings
Communication
Consistency
Coordination
Coupling
Complexity
Responsiveness/availability
Scale/elasticity
Asynchronous
Eventual
Choreographed
Very low
High
High
Very high
The Anthology Saga(aec) pattern is well suited to extremely high throughput commu‐
nication with simple or infrequent error conditions. For example, a Pipes and Filters
architecture would fit this pattern exactly.
Architects can implement the patterns described in this section in a variety of ways.
For example, architects can manage transactional sagas through atomic transactions
by using compensating updates or by managing transactional state with eventual con‐
sistency. This section showed the advantages and disadvantages of each approach,
which will help an architect decide which transactional saga pattern to use.
State Management and Eventual Consistency
State management and eventual consistency leverage finite state machines (see “Saga
State Machines” on page 352) to always know the current state of the transactional
saga, and to also eventually correct the error condition through retries or some sort
of automated or manual corrective action. To illustrate this approach, consider the
Fairy Tale Saga(seo) implementation of the ticket completion example illustrated in
Figure 12-20.
State Management and Eventual Consistency
|
351
Figure 12-20. The Fairy Tale Saga leads to better responsiveness, but leaves data sources
out of sync with one another until they can be corrected
Notice that the Survey Service is not available during the scope of the distributed
transaction. However, with this type of saga, rather than issue a compensating update,
the state of the saga is changed to NO_SURVEY and a successful response is sent to the
Sysops Expert (step 7 in the diagram). The Ticket Orchestrator Service then works
asynchronously (behind the scenes) to resolve the error programmatically by retries
and error analysis. If it cannot resolve the error, the Ticket Orchestrator Service sends
the error to an administrator or supervisor for manual repair and processing.
By managing the state of the saga rather than issuing compensating updates, the end
user (in this case, the Sysops Squad expert) doesn’t need to be concerned that the sur‐
vey was not sent to the customer—that responsibility is for the Ticket Orchestrator
Service to worry about. Responsiveness is good from the end user’s perpective, and
the user can work on other tasks while the errors are handled by the system.
Saga State Machines
A state machine is a pattern that describes all of the possible paths that can exist
within a distributed architecture. A state machine always starts with a beginning state
that launches the transactional saga, then contains transition states and correspond‐
ing action that should occur when the transition state happens.
352
|
Chapter 12: Transactional Sagas
To illustrate how a saga state machine works, consider the following workflow of a
new problem ticket created by a customer in the Sysops Squad system:
1. The customer enters a new problem ticket into the system.
2. The ticket is assigned to the next available Sysops Squad expert.
3. The ticket is then routed to the expert’s mobile device.
4. The expert receives the ticket and works on the issue.
5. The expert finishes the repair and marks the ticket as complete.
6. A survey is sent to the customer.
The various states that can exist within this transactional saga, as well as the corre‐
sponding transition actions, are illustrated in Figure 12-21. Notice that the transac‐
tional saga begins with the START node indicating the saga entry point, and terminates
with the CLOSED node indicating the saga exit point.
Figure 12-21. State diagram for creating a new problem ticket
State Management and Eventual Consistency
|
353
The following items describe in more detail this transactional saga and the corre‐
sponding states and transition actions that happen within each state:
START
The transactional saga starts with a customer entering a new problem ticket into
the system. The customer’s support plan is verified, and the ticket data is valida‐
ted. Once the ticket is inserted into the ticket table in the database, the transac‐
tional saga state moves to CREATED and the customer is notified that the ticket
has been successfully created. This is the only possible outcome for this state
transition—any errors within this state prevent the saga from starting.
CREATED
Once the ticket is successfully created, it is assigned to a Sysops Squad expert. If
no expert is available to service the ticket, it is held in a wait state until an expert
is available. Once an expert is assigned, the saga state moves to the ASSIGNED
state. This is the only outcome for this state transition, meaning the ticket is held
in CREATED state until it can be assigned.
ASSIGNED
Once a ticket is assigned to an expert, the only possible outcome is to route the
ticket to the expert. It is assumed that during the assignment algorithm, the
expert has been located and is available. If the ticket cannot be routed because the
expert cannot be located or is unavailable, the saga stays in this state until it can
be routed. Once routed, the expert must acknowledge that the ticket has been
received. Once this happens, the transactional saga state moves to ACCEPTED.
This is the only possible outcome for this state transition.
ACCEPTED
There are two possible states once a ticket has been accepted by a Sysops Squad
expert: COMPLETED or REASSIGN. Once the expert finishes the repair and
marks the ticket as “complete,” the state of the saga moves to COMPLETED.
However, if for some reason the ticket was wrongly assigned or the expert is not
able to finish the repair, the expert notifies the system and the state moves to
REASSIGN.
REASSIGN
Once in this saga state, the system will reassign the ticket to a different expert.
Like the CREATED state, if an expert is not available, the transactional saga will
remain in the REASSIGN state until an expert is assigned. Once a different expert
is found and the ticket is once again assigned, the state moves into the
ASSIGNED state, waiting to be accepted by the other expert. This is the only pos‐
sible outcome for this state transition, and the saga remains in this state until an
expert is assigned to the ticket.
354
| Chapter 12: Transactional Sagas
COMPLETED
The two possible states once an expert completes a ticket are CLOSED or
NO_SURVEY. When the ticket is in this state, a survey is sent to the customer to
rate the expert and the service, and the saga state is moved to CLOSED, thus end‐
ing the transaction saga. However, if the Survey Service is unavailable or an error
occurs while sending the survey, the state moves to NO_SURVEY, indicating that
the issue was fixed but no survey was sent to the customer.
NO_SURVEY
In this error condition state, the system continues to try sending the survey to the
customer. Once successfully sent, the state moves to CLOSED, marking the end
of the transactional saga. This is the only possible outcome of this state
transaction.
In many cases, it’s useful to put the list of all possible state transitions and the corre‐
sponding transition action in some sort of table. Developers can then use this table to
implement the state transition triggers and possible error conditions in an orchestra‐
tion service (or respective services if using choreography). An example of this prac‐
tice is shown in Table 12-10, which lists all the possible states and actions that are
triggered when the state transition occurs.
Table 12-10. Saga state machine for a new problem ticket in the Sysops Squad system
Initiating state
Transition state
Transaction action
START
CREATED
ASSIGNED
ACCEPTED
ACCEPTED
REASSIGN
COMPLETED
COMPLETED
NO_SURVEY
CREATED
ASSIGNED
ACCEPTED
COMPLETED
REASSIGN
ASSIGNED
CLOSED
NO_SURVEY
CLOSED
Assign ticket to expert
Route ticket to assigned expert
Expert fixes problem
Send customer survey
Reassign to a different expert
Route ticket to assigned expert
Ticket saga done
Send customer survey
Ticket saga done
The choice between using compensating updates or state management for distributed
transaction workflows depends on the situation as well as trade-off analysis between
responsiveness and consistency. Regardless of the technique used to manage errors
within a distributed transaction, the state of the distributed transaction should be
known and also managed.
Table 12-11 summarizes the trade-offs associated with using state management rather
than atomic distributed transactions with compensating updates.
State Management and Eventual Consistency
|
355
Trade-Offs
Table 12-11. Trade-offs associated with state management rather than atomic
distributed transactions with compensating updates
Advantages
Disadvantages
Good responsiveness
Less impact to end user for errors
Data may be out of sync when errors occur
Eventual consistency may take some time
Techniques for Managing Sagas
Distributed transactions are not something that can be simply “dropped into” a sys‐
tem. They cannot be downloaded or purchased using some sort of framework or
product like ACID transaction managers—they must be designed, coded, and main‐
tained by developers and architects.
One of the techniques we like to use to help manage distributed transactions is to lev‐
erage annotations (Java) or custom attributes (C#), or other similar artifacts in other
languages. While these language artifacts themselves don’t contain any actual func‐
tionality, they do provide a programmatic way of capturing and documenting the
transactional sagas in the system, as well as provide a means for associating services
with transactional sagas.
The source listings in Example 12-1 (Java) and Example 12-2 (C#) show an example
of implementing these annotations and custom attributes. Notice that in both imple‐
mentations, the transactional sagas (NEW_TICKET, CANCEL_TICKET, and so on) are con‐
tained within the Transaction enum, providing a single place within the source code
for listing and documenting the various sagas that exist within an application context.
Example 12-1. Source code defining a transactional saga annotation (Java)
@Retention(RetentionPolicy.RUNTIME)
@Target(ElementType.TYPE)
public @interface Saga {
public Transaction[] value();
public enum Transaction {
NEW_TICKET,
CANCEL_TICKET,
NEW_CUSTOMER,
UNSUBSCRIBE,
NEW_SUPPORT_CONTRACT
}
}
356
|
Chapter 12: Transactional Sagas
Example 12-2. Source code defining a transactional saga attribute (C#)
[AttributeUsage(AttributeTargets.Class)]
class Saga : System.Attribute {
public Transaction[] transaction;
public enum Transaction {
NEW_TICKET,
CANCEL_TICKET,
NEW_CUSTOMER,
UNSUBSCRIBE,
NEW_SUPPORT_CONTRACT
};
}
Once defined, these annotations or attributes can be used to identify services that are
involved in the transactional saga. For example, the source code listing in
Example 12-3 shows that the Survey Service (identified by the SurveyServiceAPI
class as the service entry point) is involved in the NEW_TICKET saga, whereas the
Ticket Service (identified by the TicketServiceAPI class as the service entry point) is
involved in two sagas: the NEW_TICKET and the CANCEL_TICKET.
Example 12-3. Source code showing the use of the transactional saga annotation (Java)
@ServiceEntrypoint
@Saga(Transaction.NEW_TICKET)
public class SurveyServiceAPI {
...
}
@ServiceEntrypoint
@Saga({Transaction.NEW_TICKET,)
Transaction.CANCEL_TICKET})
public class TicketServiceAPI {
...
}
Notice how the NEW_TICKET saga includes the Survey Service and the Ticket Service.
This is valuable information to a developer because it helps them define the testing
scope when making changes to a particular workflow or saga, and also lets them
know what other services might be impacted by a change to one of the services within
the transactional saga.
Techniques for Managing Sagas
|
357
Using these annotations and custom attributes, architects and developers can write
simple command-line interface (CLI) tools to walk through a codebase or source
code repository to provide saga information in real time. For example, using a simple
custom code-walk tool, a developer, architect, or even a business analyst can query
what services are involved for the NEW_TICKET saga:
$ ./sagatool.sh NEW_TICKET -services
->
->
->
->
Ticket Service
Assignment Service
Routing Service
Survey Service
$
A custom code-walking tool can look at each class file in the application context con‐
taining the @ServiceEntrypoint custom annotation (or attribute) and check the
@Saga custom annotation for the presence of the particular saga (in this case, Transac
tion.NEW_TICKET). This sort of custom tool is not complicated to write, and can help
provide valuable information when managing transactional sagas.
Sysops Squad Saga: Atomic Transactions and
Compensating Updates
Tuesday, April 5, 09:44
Addison and Austen met first thing with Logan to hash out the issues around transactionality in the new microservices architecture in the longish conference room.
Logan began, “I know that not everyone is on the same page about how what you’ve
read applies to what we’re doing here. So, I’ve prepared some workflows and diagrams
to help everyone get on the same page. Today, we’re discussing marking a ticket complete in the system. For this workflow, the Sysops Squad expert completes a job and
marks the ticket as “complete” using the mobile application on the expert’s mobile
device. I want to talk about the Epic Saga pattern and the issues around compensating updates. I’ve
created a diagram to illustrate this workflow in Figure 12-22 Can everyone see it?”
358
|
Chapter 12: Transactional Sagas
Figure 12-22. The epic saga requires the ticket status to be updated and survey to be sent
in one synchronous atomic operation
Logan continued, “I’ve also created a list that describes each step. The circled numbers on the diagram match up with the workflow.”
1. The Sysops Squad expert marks the ticket as complete using an app on their mobile device,
which is synchronously received by the Ticket Orchestrator Service.
2. The Ticket Orchestrator Service sends a synchronous request to the Ticket Service to change
the state of the ticket from “in-progress” to “complete.”
3. The Ticket Service updates the ticket number to “complete” in the database table and commits
the update.
4. As part of the ticket completion process, the Ticket Service asynchronously sends ticketing
information (such as ticket repair time, ticket wait time, duration, and so on) to a queue to be
picked up by the Analytics Service. Once sent, the Ticket Service sends an acknowledgment to
the Ticket Orchestrator Service that the update is complete.
5. At about the same time, the Analytics Service asynchronously receives the updated ticket analytics and starts to process the ticket information.
6. The Ticket Orchestrator Service then sends a synchronous request to the Survey Service to prepare and send the customer survey to the customer.
Sysops Squad Saga: Atomic Transactions and Compensating Updates
|
359
7. The Survey Service inserts data into a table with the survey information (customer, ticket info,
and timestamp) and commits the insert.
8. The Survey Service then sends the survey to the customer via email and returns an acknowledgment back to the Ticket Orchestrator Service that the survey processing is complete.
9. Finally, the Ticket Orchestrator Service sends a response back to the Sysops Squad expert’s
mobile device stating that the ticket completion processing is done. Once this happens, the
expert can select the next problem ticket assigned to them.
“Wow, this is really helpful. How long did it take you to create this?” said Addison.
“Not a little time, but it’s come in handy. You aren’t the only group that’s confused about how to get
all these moving pieces to work together. This is the hard part of software architecture. Everyone
understand the basics of the workflow?”
To a sea of nods, Logan continued, “One of the first issues that occurs with compensating updates is
that since there’s no transactional isolation within a distributed transaction (see “Distributed Transactions” on page 263), other services may have taken action on the data updated within the scope of
the distributed transaction before the distributed transaction is complete. To illustrate this issue,
consider the same Epic Saga example appearing in Figure 12-23: the Sysops Squad expert marks a
ticket as complete, but this time the Survey Service is not available. In this case, a compensating
update (step 7 in the diagram) is sent to the Ticket Service to reverse the update, changing the ticket
state from completed back to in-progress (step 8 in the diagram).”
“Notice also in Figure 12-23 that since this is an atomic distributed transaction, an error is then sent
back to the Sysops Squad expert indicating that the action was not successful and to try again. Now,
a question for you: why should the Sysops Squad expert have to worry that the survey is not sent?”
Austen pondered a moment. “But wasn’t that part of the workflow in the monolith? All that stuff
happened within a transaction, if I remember correctly.”
“Yeah, but I always thought that was weird, just never said anything,” said Addison. “I don’t see why
the expert should worry about the survey. The expert just wants to get on to the next ticket
assigned to them.”
“Right,” Logan said. “This is the issue with atomic distributed transactions—the end user is unnecessarily semantically coupled to the business process. But notice that Figure 12-23 also illustrates the
issue with the lack of transaction isolation within a distributed transaction. Notice that as part of the
original update to mark the ticket as complete, the Ticket Service asynchronously sent the ticket
information to a queue (step 4 in the diagram) to be processed by the Analytics Service (step 5).
However, when the compensating update is issued to the Ticket Service (step 7), the ticket information has already been processed by the Analytics Service in step 5.”
360
|
Chapter 12: Transactional Sagas
Figure 12-23. Epic Saga(sao) requires compensation, but side effects can occur
“We call this a side effect within distributed architectures. By reversing the transaction in the Ticket
Service, actions performed by other services using data from the prior update may have already
taken place and might not be able to be reversed. This scenario points to the importance of isolation
within a transaction, something that distributed transactions do not support. To address this issue,
the Ticket Service could send another request through the data pump to the Analytics Service, telling that service to ignore the prior ticket information, but just imagine the amount of complex code
and timing logic that would be required in the Analytics Service to address this compensating
change. Furthermore, there may have been additional downstream actions taken on the analytical
data already processed by the Analytics Service, further complicating the chain of events to reverse
and correct. With distributed architectures and distributed transactions, it really is sometimes turtles
all the way down.”
Logan paused for a moment, then continued, “Another issue—”
Austen interrupted, “Another issue?”
Logan smiled. “Another issue regarding compensating updates is compensation failures. Keeping
with the same Epic Saga example for completing a ticket, notice in Figure 12-24 that in step 7 a
compensating update is issued to the Ticket Service to change the state from completed back to inprogress. However, in this case, the Ticket Service generates an error when trying to change the state
of the ticket (step 8).”
Sysops Squad Saga: Atomic Transactions and Compensating Updates
|
361
Figure 12-24. Compensating updates within an Epic Saga can fail, leading to inconsis‐
tency and confusion about what action to take in the event of a compensation failure
“I’ve seen that happen! It took forever to track that down,” said Addison.
“Architects and developers tend to assume that compensating updates will always work,” Logan said.
“But sometimes they don’t. In this case, as shown in Figure 12-24, there is confusion about what sort
of response to send back to the end user (in this case, the Sysops Squad expert). The ticket status is
already marked as complete because the compensation failed, so attempting the “mark as complete”
request again might only lead to yet another error (such as Ticket already marked as complete). Talk
about confusion on the part of the end user!”
“Yeah, I can imagine the developers coming to us to ask us how to resolve this issue,” Addison said.
“Often developers are good checks on incomplete or confusing architecture solutions. If they are
confused, there may be a good reason,” said Logan. “OK, one more issue. Atomic distributed transactions and corresponding compensating updates also impact responsiveness. If an error occurs, the
end user must wait until all corrective action is taken (through compensating updates) before a
response is sent telling the user about the error.”
“Isn’t that where the change to eventual consistency helps, for responsiveness?” asked Austen.
362
|
Chapter 12: Transactional Sagas
“Yes, while responsiveness can sometimes be resolved by asynchronously issuing compensating
updates through eventual consistency (such as with the Parallel Saga and the Anthology Saga pattern), nevertheless most atomic distributed transactions have worse responsiveness when compensating updates are involved.”
“OK, that makes sense—atomic coordination will always have overhead,” Austen said.
“That’s a lot of information. Let’s build a table to summarize some of the trade-offs associated with
atomic distributed transactions and compensating updates.” (See Table 12-12.)
Trade-Offs
Table 12-12. Trade-offs associated with atomic distributed transactions and
compensating updates
Advantages
Disadvantages
All data restored to prior state
No transaction isolation
Allows retries and restart
Side effects may occur on compensation
Compensation may fail
Poor responsiveness for the end user
Logan said, “While this compensating transaction pattern exists, it also offers a number of challenges. Who wants to name one?”
“I know: a service cannot perform a rollback,” said Austen. “What if one of the services cannot successfully undo the previous operation? The orchestrator must have coordination code to indicate that
the transaction wasn’t successful.”
“Right—what about another?”
"To lock or not lock participating services?" said Addison. “When the mediator places a call to a service
and it updates a value, the mediator will make calls to subsequent services that are part of the workflow. However, what happens if another request appears for the first service contingent on the outcome of the first request’s resolution, either from the same mediator or a different context? This
distributed architecture problem becomes worse when the calls are asynchronous rather than synchronous (illustrated in “Phone Tag Saga(sac) Pattern” on page 330). Alternatively, the mediator could
insist that other services don’t accept calls during the course of a workflow, which guarantees a valid
transaction but destroys performance and scalability.”
Logan said, “Correct. Let’s get philosophical for a moment. Conceptually, transactions force participants to stop their individual worlds and synchronize on a particular value. This is so easy to model
with monolithic architectures and relational databases that architects overuse transactions in those
Sysops Squad Saga: Atomic Transactions and Compensating Updates
|
363
systems. Much of the real world isn’t transactional, as observed in the famous essay by Gregor
Hohpe, “Starbucks Does Not Use Two-Phase Commit”. Transactional coordination is one of the hardest parts of architecture, and the broader the scope, the worse it becomes.”
“Is there an alternative to using an Epic Saga?” Addison asked.
“Yes!” Logan said. “A more realistic approach to the scenario described in Figure 12-24 might be to
use either a Fairy Tale Saga or a Parallel Saga pattern. These sagas rely on asynchronous eventual
consistency and state management rather than atomic distributed transactions with compensating
updates when errors occur. With these types of sagas, the user is less impacted by errors that might
occur within the distributed transaction, because the error is addressed behind the scenes, without
end-user involvement. Responsiveness is also better with the state management and eventual consistency approach, because the user does not have to wait for corrective action to be taken within
the distributed transaction. If we have issues with atomicity, we can investigate those patterns as
alternatives.”
“Thanks—that’s a lot of material, but now I see why the architects made some of the decisions in the
new architecture,” Addison said.
364
| Chapter 12: Transactional Sagas
CHAPTER 13
Contracts
Friday, April 15, 12:01
Addison met with Sydney over lunch in the cafeteria to chat about coordination
between the Ticket Orchestrator and the services it integrated with for the ticket management workflow.
“Why not just use gRPC for all the communication? I heard it’s really fast,” said Sydney.
“Well, that’s an implementation, not an architecture,” Addison said. “We need to decide
what types of contracts we want before we choose how to implement them. First, we
need to decide between tight or loose contracts. Once we decide on the type, I’ll leave it to you to
decide how to implement them, as long as they pass our fitness functions.”
“What determines what kind of contract we need?” Sydney said.
In Chapter 2, we began discussing the intersection of three important forces—com‐
munication, consistency, and coordination—and how to develop trade-offs for them.
We modeled the intersectional space of the three forces in a joined three-dimensional
space, shown again in Figure 13-1. In Chapter 12, we revisited these three forces with
a discussion of the various communication styles and their trade-offs.
However much an architecture can discern a relationship like this one, some forces
cut across the conceptual space and affect all of the other dimensions equally. If pur‐
suing the visual three-dimensional metaphor, these cross-cutting forces act as an
additional dimension, much as time is orthogonal to the three physical dimensions.
365
Figure 13-1. Three-dimensional intersecting space for messaging forces in distributed
architectures
One constant factor in software architecture that cuts across and affects virtually
every aspect of architect decision making is contracts, broadly defined as how dispa‐
rate parts of an architecture connect with one another. The dictionary definition of a
contract is as follows:
contract
A written or spoken agreement, especially one concerning employment, sales, or
tenancy, that is intended to be enforceable by law.
In software, we use contracts broadly to describe things like integration points in
architecture, and many contract formats are part of the design process of software
development: SOAP, REST, gRPC, XMLRPC, and an alphabet soup of other acro‐
nyms. However, we broaden that definition and make it more consistent:
hard parts contract
The format used by parts of an architecture to convey information or
dependencies.
This definition of contract encompasses all techniques used to “wire together” parts
of a system, including transitive dependencies for frameworks and libraries, internal
and external integration points, caches, and any other communication among parts.
This chapter illustrates the effects of contracts on many parts of architecture, includ‐
ing static and dynamic quantum coupling, as well as ways to improve (or harm) the
effectiveness of workflows.
366
|
Chapter 13: Contracts
Strict Versus Loose Contracts
Like many things in software architecture, contracts don’t exist within a binary but
rather on a broad spectrum, from strict to loose. Figure 13-2 illustrates this spectrum,
using example contract types.
Figure 13-2. The spectrum of contract types, from strict to loose
A strict contract requires adherence to names, types, ordering, and all other details,
leaving no ambiguity. An example of the strictest possible contract in software is a
remote method call, using a platform mechanism such as RMI in Java. In that case,
the remote call mimics an internal method call, matching name, parameters, types,
and all other details.
Many strict contract formats mimic the semantics of method calls. For example,
developers see a host of protocols that include some variation of the RPC, tradition‐
ally an acronym for Remote Procedure Call. gRPC is an example of a popular remote
invocation framework that defaults to strict contracts.
Many architects like strict contracts because they model the identical semantic behav‐
ior of internal method calls. However, strict contracts create brittleness in integration
architecture—something to avoid. As discussed in Chapter 8, something that is
simultaneously changing frequently and used by several distinct architecture parts
creates problems in architecture. Contracts fit that description because they form the
glue within a distributed architecture: the more frequently they must change, the
more rippling problems they cause for other services. However, architects aren’t
forced to use strict contracts and should do so only when advantageous.
Even an ostensibly loose format such as JSON offers ways to selectively add schema
information to simple name-value pairs. Example 13-1 shows a strict JSON contract
with schema information attached.
Strict Versus Loose Contracts
|
367
Example 13-1. Strict JSON contract
{
"$schema": "http://json-schema.org/draft-04/schema#",
"properties": {
"acct": {"type": "number"},
"cusip": {"type": "string"},
"shares": {"type": “number", "minimum": 100}
},
"required": ["acct", "cusip", "shares"]
}
The first line references the schema definition we use and will validate against. We
define three properties (acct, cusip, and shares), along with their types and, on the
last line, which ones are required. This creates a strict contract, with required fields
and types specified.
Examples of looser contracts include formats such as REST and GraphQL, very dif‐
ferent formats but similar in demonstrating looser coupling than RPC-based formats.
For REST, the architect models resources rather than method or procedure end‐
points, making for less brittle contracts. For example, if an architect builds a RESTful
resource that describes parts of an airplane to support queries about seats, that query
won’t break in the future if someone adds details about engines to the resource—
adding more information doesn’t break what’s there.
Similarly, GraphQL is used by distributed architectures to provide read-only aggrega‐
ted data rather than perform costly orchestration calls across a wide variety of serv‐
ices. Consider the two GraphQL representations in Examples 13-2 and 13-3,
providing two different but capable views of the Profile contract.
Example 13-2. Customer Wishlist Profile representation
type Profile {
name: String
}
Example 13-3. Customer Profile representation
type Profile {
name: String
addr1: String
addr2: String
country: String
. . .
}
368
|
Chapter 13: Contracts
The concept of profile appears in both examples but with different values. In this sce‐
nario, the Customer Wishlist doesn’t have internal access to the customer’s name,
only a unique identifier. Thus, it needs access to a Customer Profile that maps the
identifier to the customer name. The Customer Profile, on the other hand, includes a
large amount of information about the customer in addition to the name. As far as
Wishlist is concerned, the only interesting thing in Profile is the name.
A common anti-pattern that some architects fall victim to is to assume that Wishlist
might eventually need all the other parts, so the architects include them in the con‐
tract from the outset. This is an example of stamp coupling and an anti-pattern in
most cases, because it introduces breaking changes where they aren’t needed, making
the architecture fragile yet providing little benefit. For example, if the Wishlist cares
about only the customer name from Profile, but the contract specifies every field in
Profile (just in case), then a change in Profile that Wishlist doesn’t care about causes a
contract breakage and coordination to fix. Keeping contracts at a “need to know”
level strikes a balance between semantic coupling and necessary information without
creating needless fragility in integration architecture.
At the far end of the spectrum of contract coupling lie extremely loose contracts,
often expressed as name-value pairs in formats like YAML or JSON, as illustrated in
Example 13-4.
Example 13-4. Name-value pairs in JSON
{
"name": "Mark",
"status": "active",
"joined": "2003"
}
Nothing but the raw facts in this example! No additional metadata, type information,
or anything else, just name-value pairs.
Using such loose contracts allows for extremely decoupled systems, often one of the
goals in architectures, such as microservices. However, the looseness of the contract
comes with trade-offs such as lack of contract certainty, verification, and increased
application logic. We illustrate in “Contracts in Microservices” on page 372 how
architects resolve this problem by using contract fitness functions.
Strict Versus Loose Contracts
|
369
Trade-Offs Between Strict and Loose Contracts
When should an architect use strict contracts and when should they use looser ones?
Like all the hard parts of architecture, no generic answer exists for this question, so it
is important for architects to understand when each is most suitable.
Strict contracts
Stricter contracts have a number of advantages, including these:
Guaranteed contact fidelity
Building schema verification within contracts ensures exact adherence to the val‐
ues, types, and other governed metadata. Some problem spaces benefit from tight
coupling for contract changes.
Versioned
Strict contracts generally require a versioning strategy to support two endpoints
that accept different values or to manage domain evolution over time. This
allows gradual changes to integration points while supporting a selective number
of past versions to make integration collaboration easier.
Easier to verify at build time
Many schema tools provide mechanisms to verify contracts at build time, adding
a level of type checking for integration points.
Better documentation
Distinct parameters and types provide excellent documentation with no
ambiguity.
Strict contracts also have a few disadvantages:
Tight coupling
By our general definition of coupling, strict contracts create tight coupling
points. If two services share a strict contract and the contract changes, both serv‐
ices must change.
Versioned
This appears in both advantages and disadvantages. While keeping distinct ver‐
sions allows for precision, it can become an integration nightmare if the team
doesn’t have a clear deprecation strategy or tries to support too many versions.
370
|
Chapter 13: Contracts
The trade-offs for strict contracts are summarized in Table 13-1.
Trade-Offs
Table 13-1. Trade-offs for strict contracts
Advantage
Disadvantage
Guaranteed contract fidelity
Versioned
Easier to verify at build time
Better documentation
Tight coupling
Versioned
Loose contracts
Loose contracts, such as name-value pairs, offer the least coupled integration points,
but they too have trade-offs, as summarized in Table 13-2.
These are some advantages of loose contracts:
Highly decoupled
Many architects have a stated goal for microservices architectures that includes
high levels of decoupling, and loose contracts provide the most flexibility.
Easier to evolve
Because little or no schema information exists, these contracts can evolve more
freely. Of course, semantic coupling changes still require coordination across all
interested parties—implementation cannot reduce semantic coupling—but loose
contracts allow easier implementation evolution.
Loose contracts also have a few disadvantages:
Contract management
Loose contracts by definition don’t have strict contract features, which may cause
problems such as misspelled names, missing name-value pairs, and other defi‐
ciencies that schemas would fix.
Requires fitness functions
To solve the contract issues just described, many teams use consumer-driven
contracts as an architecture fitness function to make sure that loose contracts still
contain sufficient information for the contract to function.
Strict Versus Loose Contracts
|
371
Trade-Offs
Table 13-2. Trade-offs for loose contracts
Advantage
Disadvantage
Highly decoupled
Easier to evolve
Contract management
Requires fitness functions
For an example of the common trade-offs encountered by architects, consider the
example of contracts in microservice architectures.
Contracts in Microservices
Architects must constantly make decisions about how services interact with one
another, what information to pass (the semantics), how to pass it (the implementa‐
tion), and how tightly to couple the services.
Coupling levels
Consider two microservices with independent transactionality that must share
domain information such as Customer Address, shown in Figure 13-3.
Figure 13-3. Two services that must share domain information about the customer
The architect could implement both services in the same technology stack and use a
strictly typed contract, either a platform-specific remote procedure protocol (such as
RMI) or an implementation-independent one like gRPC, and pass the customer
information from one to another with high confidence of contract fidelity. However,
this tight coupling violates one of the aspirational goals of microservices architec‐
tures, where architects try to create decoupled services.
372
|
Chapter 13: Contracts
Consider the alternative approach, where each service has its own internal represen‐
tation of Customer, and the integration uses name-value pairs to pass information
from one service to another, as illustrated in Figure 13-4.
Here, each service has its own bounded-context definition of Customer. When pass‐
ing information, the architect utilizes name-value pairs in JSON to pass the relevant
information in a loose contract.
Figure 13-4. Microservices with their own internal semantic representation can pass
values in simple messages
This loose coupling satisfies many of the overarching goals of microservices. First, it
creates highly decoupled services modeled after bounded contexts, allowing each
team to evolve internal representations as aggressively as needed. Second, it creates
implementation decoupling. If both services start in the same technology stack, but
the team in the second decides to move to another platform, it likely won’t affect the
first service at all. All platforms in common use can produce and consume namevalue pairs, making them the lingua franca of integration architecture.
The biggest downside of loose contracts is contract fidelity—as an architect, how can
I know that developers pass the correct number and type of parameters for integra‐
tion calls? Some protocols, such as JSON, include schema tools to allow architects to
overlay loose contracts with more metadata. Architects can also use a style of archi‐
tect fitness function called a consumer-driven contract.
Consumer-driven contracts
A common problem in microservices architectures is the seemingly contradictory
goals of loose coupling yet contract fidelity. One innovative approach that utilizes
advances in software development is a consumer-driven contract, common in micro‐
services architectures.
In many architecture integration scenarios, a service decides what information to
emit to other integration partners (a push model—the service provider pushes a con‐
Strict Versus Loose Contracts
|
373
tract to consumers). The concept of a consumer-driven contract inverses that rela‐
tionship into a pull model; here, the consumer puts together a contract for the items
they need from the provider, and passes the contract to the provider, who includes it
in their build and keeps the contract test green at all times. The contract encapsulates
the information the consumer needs from the provider. This may work for a network
of interlocking requests that the Provider must honor, as illustrated in Figure 13-5.
Figure 13-5. Consumer-driven contracts allow the provider and consumers to stay in
sync via automated architectural governance
In this example, the team on the left provides bits of (likely) overlapping information
to each of the consumer teams on the right. Each consumer creates a contract specify‐
ing required information and passes it to the provider, who includes their tests as part
of a continuous integration or deployment pipeline. This allows each team to specify
the contract as strictly or loosely as needed while guaranteeing contract fidelity as
part of the build process. Many consumer-driven contract testing tools provide facili‐
ties to automate build-time checks of contracts, providing another layer of benefit
similar to stricter contracts.
Consumer-driven contracts are quite common in microservices architecture because
they allow architects to solve the dual problems of loose coupling and governed inte‐
gration. Trade-offs of consumer-driven contracts are shown in Table 13-3.
Advantages of consumer-driven contracts are as follows:
Allow loose contract coupling between services
Using name-value pairs is the loosest possible coupling between two services,
allowing implementation changes with the least chance of breakage.
374
|
Chapter 13: Contracts
Allow variability in strictness
If teams use architecture fitness functions, architects can build stricter verifica‐
tions than typically offered by schemas or other type-additive tools. For example,
most schemas allow architects to specify things like numeric type but not accept‐
able ranges of values. Building fitness functions allows architects to build as
much specificity as they like.
Evolvable
Loose coupling implies evolvability. Using simple name-value pairs allows inte‐
gration points to change implementation details without breaking the semantics
of the information passed between services.
These are disadvantages of consumer-driven contracts:
Require engineering maturity
Architecture fitness functions are a great example of a capability that really works
well only when well-disciplined teams have good practices and don’t skip steps.
For example, if all teams run continuous integration that includes contract tests,
then fitness functions provide a good verification mechanism. On the other
hand, if many teams ignore failed tests or are not timely in running contract
tests, integration points may be broken in architecture longer than desired.
Two interlocking mechanisms rather than one
Architects often look for a single mechanism to solve problems, and many of the
schema tools have elaborate capabilities to create end-to-end connectivity. How‐
ever, sometimes two simple interlocking mechanisms can solve the problem
more simply. Thus, many architects use the combination of name-value pairs and
consumer-driven contracts to validate contracts. However, this means that teams
require two mechanisms rather than one.
The architect’s best solution for this trade-off comes down to team maturity and
decoupling with loose contracts versus complexity plus certainty with stricter
contracts.
Trade-Offs
Table 13-3. Trade-offs for consumer-driven contracts
Advantage
Disadvantage
Allows loose contract coupling between services Requires engineering maturity
Allows variability in strictness
Two interlocking mechanisms rather than one
Evolvable
Strict Versus Loose Contracts
|
375
Stamp Coupling
A common pattern and sometimes anti-pattern in distributed architectures is stamp
coupling, which describes passing a large data structure between services, but each
service interacts with only a small part of the data structure. Consider the example of
four services shown in Figure 13-6.
Figure 13-6. Stamp coupling between four services
Each service accesses (either reads, writes, or both) only a small portion of the data
structure passed between each service. This pattern is common when an industrystandard document format exists, typically in XML. For example, the travel industry
has a global standard XML document format that specifies details about travel itin‐
eraries. Several systems that work with travel-related services pass the entire docu‐
ment around, updating only their relevant sections.
Stamp coupling, however, is often an accidental anti-pattern, where an architect has
over-specified the details in a contract that aren’t needed or accidentally consumes far
too much bandwidth for mundane calls.
Over-Coupling via Stamp Coupling
Going back to our Wishlist and Profile Services, consider tying the two together with
a strict contract combined with stamp coupling, as illustrated in Figure 13-7.
In this example, even though the Wishlist Service needs only the name (accessed via a
unique ID), the architect has coupled Profile’s entire data structure as the contract,
perhaps in a misguided effort for future proofing. However, the negative side effect of
too much coupling in contracts is brittleness. If Profile changes a field that Wishlist
doesn’t care about, such as state, it still breaks the contract.
376
|
Chapter 13: Contracts
Figure 13-7. The Wishlist Service is stamp coupled to the Profile Service
Over-specifying details in contracts is generally an anti-pattern but easy to fall into
when also using stamp coupling for legitimate concerns, including uses such as work‐
flow management (see “Stamp Coupling for Workflow Management” on page 378).
Bandwidth
The other inadvertent anti-pattern that some architects fall into is one of the famous
fallacies of distributed computing: bandwidth is infinite. Architects and developers
rarely have to consider the cumulative size of the number of method calls they make
within a monolith because natural barriers exist. However, many of those barriers dis‐
appear in distributed architectures, inadvertently creating problems.
Consider the previous example for 2,000 requests per second. If each payload is 500
KB, then the bandwidth required for this single request equals 1,000,000 KB per sec‐
ond! This is obviously an egregious use of bandwidth for no good reason. Alterna‐
tively, if the coupling between Wishlist and Profile contained only the necessary
information, name, the overhead changes to 200 bytes per second, for a perfectly rea‐
sonable 400 KB.
Stamp coupling can create problems when overused, including issues caused by cou‐
pling too tightly to bandwidth. However, like all things in architecture, it has benefi‐
cial uses as well.
Stamp Coupling
|
377
Stamp Coupling for Workflow Management
In Chapter 12, we covered a number of dynamic quantum communication patterns,
including several that featured the coordination style of choreography. Architects tend
toward mediation for complex workflows for the many reasons we’ve delineated.
However, what if other factors, such as scalability, drive an architect toward a solution
that is both choreographed and complex?
Architects can use stamp coupling to manage the workflow state between services,
passing both domain knowledge and workflow state as part of the contract, as illus‐
trated in Figure 13-8.
Figure 13-8. Using stamp coupling for workflow management
In this example, an architect designs the contract to include workflow information:
status of the workflow, transactional state, and so on. As each domain service accepts
the contract, it updates its portion of the contract and state for the workflow, then
passes it along. At the end of the workflow, the receiver can query the contract to
determine success or failure, along with status and information such as error mes‐
sages. If the system needs to implement transactional consistency throughout, then
domain services should rebroadcast the contract to previously visited services to
restore atomic consistency.
Using stamp coupling to manage workflow does create higher coupling between serv‐
ices than nominal, but the semantic coupling must go somewhere—remember, an
architect cannot reduce semantic coupling via implementation. However, in many
cases, switching to choreography can improve throughput and scalability, making the
choice of stamp coupling over mediation an attractive one. Table 13-4 shows the
trade-offs for stamp coupling.
378
|
Chapter 13: Contracts
Trade-Offs
Table 13-4. Trade-offs for stamp coupling
Advantage
Disadvantage
Allows complex workflows within
choreographed solutions
Creates (sometimes artificially) high coupling between
collaborators
Can create bandwidth issues at high scale
Sysops Squad Saga: Managing Ticketing Contracts
Tuesday, May 10, 10:10
Sydney and Addison met again in the cafeteria over coffee to discuss the contracts in
the ticket management workflow.
Addison said, “Let’s look at the workflow under discussion, the ticket management workflow. I’ve sketched out the types of contracts we should use, and wanted to run it by you
to make sure I wasn’t missing anything. It’s illustrated in Figure 13-9.”
“The contracts between the orchestrator and the two ticket services, Ticket Management and Ticket Assignment, are tight; that information is highly semantically coupled and likely to
change together,” Addison said. “For example, if we add new types of things to manage, the assignment must sync up. The Notification and Survey Service can be much looser—the information
changes more slowly, and doesn’t benefit from brittle coupling.”
Sydney said, “All those decisions make sense—but what about the contract between the orchestrator and the Sysops Squad expert application? It seems that would need as tight a contract as
assignment.”
“Good catch—nominally, we would like the contract with the mobile application to match ticket
assignment. However, we deploy the mobile application through a public app store, and their
approval process sometimes takes a long time. If we keep the contracts looser, we gain flexibility
and slower rate of change.”
Sysops Squad Saga: Managing Ticketing Contracts
|
379
Figure 13-9. Types of contracts between collaborators in the ticket management
workflow
They both wrote an ADR for this:
ADR: Loose Contract for Sysops Squad Expert Mobile Application
Context
The mobile application used by Sysops Squad experts must be deployed through the
public app store, imposing delays on the ability to update contracts.
Decision
We will use a loose, name-value pair contract to pass information to and from the orchestrator and the mobile application.
We will build an extension mechanism to allow temporary extensions for short-term
flexibility.
Consequences
The decision should be revisited if the app store policy allows for faster (or continuous)
deployment.
More logic to validate contracts must reside in the orchestrator and mobile application.
380
|
Chapter 13: Contracts
CHAPTER 14
Managing Analytical Data
Tuesday, May 31, 13:23
Logan and Dana (the data architect) were standing outside the big conference room,
chatting after the weekly status meeting.
“How are we going to handle analytical data in this new architecture?” asked Dana.
“We’re splitting the databases into small parts, but we’re going to have to glue all that
data back together for reporting and analytics. One of the improvements we’re trying to
implement is better predictive planning, which means we are using more data science
and statistics to make more strategic decisions. We now have a team that thinks about
analytical data, and we need a part of the system to handle this need. Are we going to have a data
warehouse?”
Logan said, “We looked into creating a data warehouse, and while it solved the consolidation problem, it had a bunch of issues for us.”
Much of this book has been concerned with how to analyze trade-offs within existing
architectural styles such as microservices. However, the techniques we highlight can
also be used to understand brand-new capabilities as they appear in the software
development ecosystem; data mesh is an excellent example.
Analytical and operational data have widely different purposes in modern architec‐
tures (see “The Importance of Data in Architecture” on page 4); much of this book
has dealt with the difficult trade-offs associated with operational data. When client/
server systems became popular and powerful enough for large enterprises, architects
and database administrators looked for a solution that would allow specialized
queries.
381
Previous Approaches
The split between operational and analytical data is hardly a new problem—the fun‐
damental different uses of data have existed as long as data. As architecture styles
have emerged and evolved, approaches for how to handle data have changed and
evolved similarly.
The Data Warehouse
Back in earlier eras of software development (for example, mainframe computers or
early personal computers), applications were monolithic, including code and data on
the same physical system. Not surprisingly, given the context we’ve covered up until
this point, transaction coordination across different physical systems became chal‐
lenging. As data requirements became more ambitious, coupled with the advent of
local area networks in offices, this led to the rise of client/server applications, where a
powerful database server runs on the network and desktop applications run on local
computers, accessing data over the network. The separation of application and data
processing allowed better transactional management, coordination, and numerous
other benefits, including the ability to start utilizing historical data for new purposes,
such as analytics.
Architects made an early attempt to provide queriable analytical data with the Data
Warehouse pattern. The basic problem they tried to address goes to the core of the
separation of operational and analytical data: the formats and schemas of one don’t
necessarily fit (or even allow the use of) the other. For example, many analytical
problems require aggregations and calculations, which are expensive operations on
relational databases, especially those already operating under heavy transactional
load.
The Data Warehouse patterns that evolved had slight variations, mostly based on ven‐
dor offerings and capabilities. However, the pattern had many common characteris‐
tics. The basic assumption was that operational data was stored in relational databases
directly accessible via the network. Here are the main characteristics of the Data
Warehouse pattern:
Data extracted from many sources
As the operational data resided in individual databases, part of this pattern speci‐
fied a mechanism for extracting the data into another (massive) data store, the
“warehouse” part of the pattern. It wasn’t practical to query across all the various
databases in the organization to build reports, so the data was extracted into the
warehouse solely for analytical purposes.
382
|
Chapter 14: Managing Analytical Data
Transformed to single schema
Often, operational schemas don’t match the ones needed for reporting. For
example, an operational system needs to structure schemas and behavior around
transactions, whereas an analytical system is rarely OLTP data (see Chapter 1)
but typically deals with large amounts of data, for reporting, aggregations, and so
on. Thus, most data warehouses utilized a Star Schema to implement dimensional
modelling, transforming data from operational systems in differing formats into
the warehouse schema. To facilitate speed and simplicity, warehouse designers
denormalize the data to facilitate performance and simpler queries.
Loaded into warehouse
Because the operational data resides in individual systems, the warehouse must
build mechanisms to regularly extract the data, transform it, and place it in the
warehouse. Designers either used built-in relational database mechanisms like
replication or specialized tools to build translators from the original schema to
the warehouse schema. Of course, any changes to operational systems schemas
must be replicated in the transformed schema, making change coordination
difficult.
Analysis done on the warehouse
Because the data “lives” in the warehouse, all analysis is done there. This is desir‐
able from an operational standpoint: the data warehouse machinery typically fea‐
tured massively capable storage and compute, offloading the heavy requirements
into its own ecosystem.
Used by data analysts
The data warehouse utilized data analysts, whose job included building reports
and other business intelligence assets. However, building useful reports requires
domain understanding, meaning that domain expertise must reside in both the
operational data system and the analytical systems, where query designers must
use the same data in a transformed schema to build meaningful reports and busi‐
ness intelligence.
BI reports and dashboards
The output of the data warehouse included business intelligence reports, dash‐
boards that provide analytical data, reports, and any other information to allow
the company to make better decisions.
SQL-ish interface
To make it easier for DBAs to use, most data warehouse query tools provided
familiar affordances, such as a SQL-like language for forming queries. One of the
reasons for the data transformation step mentioned previously was to provide
users with a simpler way to query complex aggregations and other intelligence.
Previous Approaches
|
383
The Star Schema
The Star Schema pattern was popular with data marts and warehouses. It separates the
data semantics into facts, which hold the organization’s quantifiable data, and dimen‐
sions; hence they are also known as dimensional models, which include descriptive
attributes of the fact data.
Examples of fact data for the Sysops Squad might include hourly rate, time to repair,
distance to client, and other concretely measurable things. Dimensions might include
squad member specialties, squad person names, store locations, and other metadata.
Most significantly, the Star Schema is purposely denormalized to facilitate simpler
queries, simplified business logic (in other words, fewer complex joins), faster queries
and aggregations, complex analytics such as data cubes, and the ability to form muti‐
dimensional queries. Most Star Schemas become incredibly complex.
The Data Warehouse pattern provides a good example of technical partitioning in
software architecture: warehouse designers transform the data into a schema that
facilitates queries and analysis but loses any domain partitioning, which must be recreated in queries where required. Thus, highly trained specialists were required to
understand how to construct queries in this architecture.
However, the major failings of the Data Warehouse pattern included integration brit‐
tleness, extreme partitioning of domain knowledge, complexity, and limited function‐
ality for intended purpose:
Integration brittleness
The requirement built into this pattern to transform the data during the injection
phase creates crippling brittleness in systems. A database schema for a particular
problem domain is highly coupled to the semantics of that problem; changes to
the domain require schema changes, which in turn require data import logic
changes.
Extreme partitioning of domain knowledge
Building complex business workflows requires domain knowledge. Building
complex reports and business intelligence also requires domain knowledge, cou‐
pled with specialized analytics techniques. Thus, the Venn diagrams of domain
expertise overlap, but only partially. Architects, developers, DBAs, and data sci‐
entists must all coordinate on data changes and evolution, forcing tight coupling
between vastly different parts of the ecosystem.
384
| Chapter 14: Managing Analytical Data
Complexity
Building an alternate schema to allow advanced analytics adds complexity to the
system, along with the ongoing mechanisms required to injest and transform
data. A data warehouse is a separate project outside the normal operational
systems for an organization, so must be maintained as a wholly separate ecosys‐
tem, yet highly coupled to the domains embedded inside the operational systems.
All these factors contribute to complexity.
Limited functionality for intended purpose
Ultimately, most data warehouses failed because they didn’t deliver business value
commensurate to the effort required to create and maintain the warehouse.
Because this pattern was common long before cloud environments, the physical
investment in infrastructure was huge, along with the ongoing development and
maintenance. Often, data consumers would request a certain type of report that
the warehouse couldn’t provide. Thus, such an ongoing investment for ultimately
limited functionality doomed most of these projects.
Synchronization creates bottlenecks
The need in a data warehouse to synchronize data across a wide variety of opera‐
tional systems creates both operational and organizational bottlenecks—a loca‐
tion where multiple and otherwise independent data streams must converge. A
common side effect of the data warehouse is the synchronization process impact‐
ing operational systems despite the desire for decoupling.
Operational versus analytical contract differences
Systems of record have specific contract needs (discussed in Chapter 13). Analyt‐
ical systems also have contractual needs that often differ from the operational
ones. In a data warehouse, the pipelines often handle the transformation as well
as ingestion, introducing contractual brittleness in the transformation process.
Table 14-1 shows the trade-offs for the data warehouse pattern.
Trade-Offs
Table 14-1. Trade-offs for the Data Warehouse pattern
Advantage
Disadvantage
Centralized consolidation of data
Dedicated analytics silo provides isolation
Extreme partitioning of domain knowledge
Integration brittleness
Complexity
Limited functionality for intended purpose
Previous Approaches
|
385
Tuesday, May 31, 13:33
“We looked at creating a data warehouse, but realized that it fit better with older, monolithic kinds of architectures than modern distributed ones,” said Logan. “Plus, we have a
ton more machine learning cases now that we need to support.”
“What about the data lake idea I’ve been hearing about?” asked Dana. “I read a blog post
on Martin Fowler’s site.1 It seems like it addresses a bunch of the issues with the data
warehouse, and it is more suitable for ML use cases.”
“Oh, yes, I read that post when it came out,” Logan said. “His site is a treasure trove of good information, and that post came out right after the topic of microservices became hot. In fact, I first read
about microservices on that same site in 2014, and one of the big questions at the time was, How do
we manage reporting in architectures like that? The data lake was one of the early answers, mostly as a
counter to the data warehouse, which definitely won’t work in something like microservices.”
“Why not?” Dana asked.
The Data Lake
As in many reactionary responses to the complexity, expense, and failures of the data
warehouse, the design pendulum swung to the opposite pole, exemplified by the Data
Lake pattern, intentionally the inverse of the Data Warehouse pattern. While it keeps
the centralized model and pipelines, it inverts the “transform and load” model of the
data warehouse to a “load and transform” one. Rather than do the immense work of
transformation, the philosophy of the Data Lake pattern holds that, rather than do
useless transformations that may never be used, do no transformations, allowing
business users access to analytical data in its natural format, which typically required
transformation and massaging for their purpose. Thus, the burden of work was made
reactive rather than proactive—rather than do work that might not be needed, do
transformation work only on demand.
The basic observation that many architects made was that the prebuilt schemas in
data warehouses were frequently not suited to the type of report or inquiry required
by users, requiring extra work to understand the warehouse schema enough to craft a
solution. Additionally, many machine learning models work better with data “closer”
to the semi-raw format rather than a transformed version. For domain experts who
already understood the domain, this presented an excruciating ordeal, where data
was stripped of domain separation and context to be transformed into the data
1 Martin Fowler posted an influential message about the Data Lake pattern on his blog in 2015 at https://martin
fowler.com/bliki/DataLake.html.
386
|
Chapter 14: Managing Analytical Data
warehouse, only to require domain knowledge to craft queries that weren’t natural fits
of the new schema!
Characteristics of the Data Lake pattern are as follows:
Data extracted from many sources
Operational data is still extracted in this pattern, but less transformation into
another schema takes place—rather, the data is often stored in its “raw,” or native,
form. Some transformation may still occur in this pattern. For example, an
upstream system might dump formatted files into a lake that are organized based
on a column-based snapshots.
Loaded into the lake
The lake, often deployed in cloud environments, consists of regular data dumps
from the operational systems.
Used by data scientists
Data scientists and other consumers of analytical data discover the data in the
lake and perform whatever aggregations, compositions, and other transforma‐
tions necessary to answer specific questions.
The Data Lake pattern, while an improvement in many ways to the Data Warehouse
pattern, still suffered many limitations.
This pattern still takes a centralized view of data, where data is extracted from opera‐
tional systems’ databases and replicated into a more or less free-form lake. The bur‐
den was on the consumer to discover how to connect disparate data sets together,
which often happened in the data warehouse despite the level of planning. The logic
followed that, if we’re going to have to do pre-work for some analytics, let’s do it for
all, and skip the massive up-front investment.
While the Data Lake pattern avoided the transformation-induced problems from the
Data Warehouse pattern, it also either didn’t address or created new problems.
Difficulty in discovery of proper assets
Much of the understanding of data relationships within a domain evaporates as
data flows into the unstructured lake. Thus, domain experts must still involve
themselves in crafting analysis.
PII and other sensitive data
Concern around PII has risen in concert with the capabilities of the data scientist
to take disparate pieces of information and learn privacy-invading knowledge.
Many countries now restrict not just private information, but also information
that can be combined to learn and identify, for ad targeting or other less savory
purposes. Dumping unstructured data into a lake often risks exposing informa‐
tion that can be stitched together to violate privacy. Unfortunately, just as in the
Previous Approaches
|
387
discovery process, domain experts have the knowledge necessary to avoid acci‐
dental exposures, forcing them to reanalyze data in the lake.
Still technically, not domain, partitioned
The current trend in software architecture shifts focus from partitioning a system
based on technical capabilities into ones based on domains, whereas both the
Data Warehouse and Data Lake patterns focus on technical partitioning. Gener‐
ally, architects design each of those solutions with distinct ingestion, transforma‐
tion, loading, and serving partitions, each focused on a technical capability.
Modern architecture patterns favor domain partitioning, encapsulating technical
implementation details. For example, the microservices architecture attempts to
separate services by domain rather than technical capabilities, encapsulating
domain knowledge, including data, inside the service boundary. However, both
the Data Warehouse and Data Lake patterns try to separate data as a separate
entity, losing or obscuring important domain perspectives (such as PII data) in
the process.
The last point is critical—increasingly, architects design around domain rather than
technical partitioning in architecture, and both previous approaches exemplify sepa‐
rating data from its context. What architects and data scientists need is a technique
that preserves the appropriate kind of macro-level partitioning, yet supports a clean
separation of analytical from operational data. Table 14-2 lists the trade-offs for the
Data Lake pattern.
Trade-Offs
Table 14-2. Trade-offs for the Data Lake pattern
Advantage
Disadvantage
Less structured than data warehouse
Less up-front transformation
Better suited to distributed architectures
Sometimes difficult to understand relationships
Requires ad hoc transformations
The disadvantages around brittleness and pathological coupling of pipelines remain.
Although they do less transformation in the Data Lake pattern, it is still common, as
well as data cleansing.
The Data Lake pattern pushes data integrity testing, data quality, and other quality
issues to downstream lake pipelines, which can create some of the same operational
bottlenecks that manifest in the Data Warehouse pattern.
388
|
Chapter 14: Managing Analytical Data
Because of both technical partitioning and the batch-like nature, solutions may suffer
from data staleness. Without careful coordination, architects either ignore the
changes in upstream systems, resulting in stale data, or allow the coupled pipelines to
break.
Tuesday, May 31, 14:43
“OK, so we can’t use the data lake either!” exclaimed Dana. “What now?”
“Fortunately, some recent research has found a way to solve the problem of analytical
data with distributed architectures like microservices,” replied Logan. “It adheres to the
domain boundaries we’re trying to achieve, but also allows us to project analytical data
in a way that the data scientists can use. And, it eliminates the PII problems our lawyers
are worried about.”
“Great!” Dana replied. “How does it work?”
The Data Mesh
Observing other trends in distributed architectures, Zhamak Dehghani and several
other innovators derived the core idea of the Data Mesh pattern from domainoriented decoupling of microservices, service mesh, and sidecars (see “Sidecars and
Service Mesh” on page 234), and applied it to analytical data, with modifications. As
we mentioned in Chapter 8, the Sidecar Pattern provides a nonentangling way to
organize orthogonal coupling (see “Orthogonal Coupling” on page 238); the separa‐
tion between operational and analytical data is another excellent example of just such
a coupling, but with more complexity than simple operational coupling.
Definition of Data Mesh
Data mesh is a sociotechnical approach to sharing, accessing, and managing analyti‐
cal data in a decentralized fashion. It satisfies a wide range of analytical use cases,
such as reporting, ML model training, and generating insights. Contrary to the previ‐
ous architecture, it does so by aligning the architecture and ownership of the data
with the business domains and enabling a peer-to-peer consumption of data.
Data mesh is founded on four principles:
Domain ownership of data
Data is owned and shared by the domains that are most intimately familiar with
the data: the domains that either are originating the data, or are the first-class
consumers of the data. This architecture allows for distributed sharing and
accessing the data from multiple domains and in a peer-to-peer fashion without
The Data Mesh
|
389
any intermediary and centralized lake or warehouse, and without a dedicated
data team.
Data as a product
To prevent siloing of data and encourage domains to share their data, data mesh
introduces the concept of data served as a product. It puts in place the organiza‐
tional roles and success metrics necessary to ensure that domains provide their
data in a way that delights the experience of data consumers across the organiza‐
tion. This principle leads to the introduction of a new architectural quantum
called data product quantum, to maintain and serve discoverable, understanda‐
ble, timely, secure, and high-quality data to the consumers. This chapter introdu‐
ces the architectural aspect of the data product quantum.
Self-serve data platform
To empower the domain teams to build and maintain their data products, data
mesh introduces a new set of self-serve platform capabilities. The capabilities
focus on improving the experience of data product developers and consumers. It
includes features such as declarative creation of data products, discoverability of
data products across the mesh through search and browsing, and managing the
emergence of other intelligent graphs, such as lineage of data and knowledge
graphs.
Computational federated governance
This principle assures that despite decentralized ownership of the data,
organization-wide governance requirements—such as compliance, security, pri‐
vacy, and quality of data, as well as interoperability of data products—are met
consistently across all domains. Data mesh introduces a federated decisionmaking model composed of domain data product owners. The policies they for‐
mulate are automated and embedded as code in each and every data product.
The architectural implication of this approach to governance is a platformsupplied embedded sidecar in each data product quantum to store and execute
the policies at the point of access: data read or write.
Data mesh is a wide-ranging topic, fully covered in the book Data Mesh by Zhamak
Dehghani (O’Reilly). In this chapter, we focus on the core architectural element, the
data product quantum.
Data Product Quantum
The core tenet of the data mesh overlays modern distributed architectures such as
microservices. Just as in the service mesh, teams build a data product quantum (DPQ)
adjacent but coupled to their service, as illustrated in Figure 14-1.
390
|
Chapter 14: Managing Analytical Data
Figure 14-1. Structure of a data product quantum
In this example, the service Alpha contains both behavior and transactional (opera‐
tional) data. The domain includes a data product quantum, which also contains code
and data, and which acts as an interface to the overall analytical and reporting por‐
tion of the system. The DPQ acts as an operationally independent but highly coupled
set of behaviors and data.
Several types of DPQs commonly exist in modern architectures:
Source-aligned (native) DPQ
Provides analytical data on behalf of the collaborating architecture quantum, typ‐
ically a microservice, acting as a cooperative quantum.
Aggregate DQP
Aggregates data from multiple inputs, either synchronously or asynchronously.
For example, for some aggregations, an asynchronous request may be sufficient;
for others, the aggregator DPQ may need to perform synchronous queries for a
source-aligned DPQ.
The Data Mesh
|
391
Fit-for-purpose DPQ
A custom-made DPQ to serve a particular requirement, which may encompass
analytical reporting, business intelligence, machine learning, or other supporting
capability.
Each domain that also contributes to analysis and business intelligence includes a
DPQ, as illustrated in Figure 14-2.
Figure 14-2. The data product quantum acts as a separate but highly coupled adjunct to
a service
Here, the DPQ represents a component owned by the domain team responsible for
implementing the service. It overlaps information stored in the database, and may
have interactions with some of the domain behavior asynchronously. The data
392
| Chapter 14: Managing Analytical Data
product quantum also likely has behavior as well as data for the purposes of analytics
and business intelligence.
Each data product quantum acts as a cooperative quantum for the service itself:
Cooperative quantum
An operationally separate quantum that communicates with its cooperator via
asynchronous communication and eventual consistency, yet features tight con‐
tract coupling with its cooperator and generally looser contract coupling to the
analytics quantum, the service responsible for reports, analysis, business intelli‐
gence, and so on. While the two cooperating quanta are operationally independ‐
ent, they represent two sides of data: operational data in the quantum and
analytical data in the data product quantum.
Some portion of the system will carry the responsibility for analytics and business
intelligence, which will form its own domain and quantum. To operate, this analytical
quantum has static quantum coupling to the individual data product quanta it needs
for information. This service may make either synchronous or asynchronous calls to
the DPQ, depending on the type of request. For example, some DPQs will feature a
SQL interface to the analytical DPQ, allowing synchronous queries. Other require‐
ments may aggregate information across multiple DPQs.
Data Mesh, Coupling, and Architecture Quantum
Because analytical reporting is probably a required feature of a solution, the DPQ and
its communication implementation belong to the static coupling of an architecture
quantum. For example, in a microservices architecture, the service plane must be
available, just as a message broker must be available if the design calls for messaging.
However, like the Sidecar pattern in a service mesh, the DPQ should be orthogonal to
implementation changes within the service, and maintain a separate contract with the
data plane.
From a dynamic quantum coupling standpoint, the data sidecar should always imple‐
ment one of the communication patterns that features both eventual consistency and
asynchronicity: either the “Parallel Saga(aeo) Pattern” on page 346 or “Anthology
Saga(aec) Pattern” on page 349. In other words, a data sidecar should never include a
transactional requirement to keep operational and analytical data in sync, which
would defeat the purpose of using a DPQ for orthogonal decoupling. Similarly, com‐
munication to the data plane should genearlly be asynchronous, so as to have mini‐
mal impact on the operational architecture characteristics of the domain service.
When to Use Data Mesh
Like all things in architecture, this pattern has trade-offs associated with it, as shown
in Table 14-3.
The Data Mesh
|
393
Trade-Offs
Table 14-3. Trade-offs for the Data Mesh pattern
Advantage
Disadvantage
Highly suitable for microservices architectures
Requires contract coordination
with data product quantum
Requires asynchronous
communication and eventual
consistency
Follows modern architecture principles and engineering
practices
Allows excellent decoupling between analytical and
operational data
Carefully formed contracts allow loosely coupled evolution
of analytical capabilities
It is most suitable in modern distributed architectures such as microservices with
well-contained transactionality and good isolation between services. It allows domain
teams to determine the amount, cadence, quality, and transparency of the data con‐
sumed by other quanta.
It is more difficult in architectures where analytical and operational data must stay in
sync at all times, which presents a daunting challenge in distributed architectures.
Finding ways to support eventual consistency, perhaps with very strict contracts,
allows many patterns that don’t impose other difficulties.
Data mesh is an outstanding example of the constant incremental evolution that
occurs in the software development ecosystem; new capabilities create new perspec‐
tives, which in turn help address some persistent headaches from the past, such as the
artificial separation of domain from data, both operational and analytical.
Sysops Squad Saga: Data Mesh
Friday, June 10, 09:55
Logan, Dana, and Addison met in the big conference room, which often had leftover
snacks (or, this early in the day, breakfast) from previous meetings.
“I just returned from a meeting with our data scientists, and they are trying to figure out
a way we can solve a long-term problem for us—we need to become data-driven in
expert supply planning, for skill sets demand for different geographical locations at different points in time. That capability will help recruitment, training, and other supplyrelated functions,” said Logan.
394
|
Chapter 14: Managing Analytical Data
“I haven’t been involved in much of the data mesh implementation—how far along are we?” asked
Addison.
“Each new service we’ve implemented includes a DPQ. The domain team is responsible for running
and maintaining the DQP cooperative quantum for their service. We’ve only just started. We’re gradually building out the capabilities as we identify the needs. I have a picture of the Ticket Management Domain in Figure 14-3.”
Figure 14-3. Ticket Management Domain, including two services with their own DPQs,
with a Tickets DPQ
Logan said, “Tickets DPQ is its own architecture quantum, and acts as an aggregation point for a
couple of different ticket views that other systems care about.”
“How much does each team have to build versus already supplied?” Addison asked.
“I can answer that,” said Dana. “The data mesh platform team is supplying the data users and data
product developers with a set of self-serve capabilities. That allows any team that wants to build a
new analytical use case to search and find the data products of choice within existing architecture
quanta, directly connect to them, and start using them. The platform also supports domains that
want to create new data products. The platform continuously monitors the mesh for any data product downtimes, or incompatibility with the governance policies and informs the domain teams to
take actions.”
Logan said, “The domain data product owners in collaboration with security, legal, risk, and compliance SMEs, as well as the platform product owners, have formed a global federated governance
group, which decides on aspects of the DPQs that must be standardized, such as their data-sharing
contracts, modes of asynchronous transport of data, access control, and so on. The platform team,
over a span of time, enriches the DPQ’s sidecar with new policy execution capabilities and upgrades
the sidecars uniformly across the mesh.”
Sysops Squad Saga: Data Mesh
|
395
“Wow, we’re further along that I thought,” said Dana. “What data do we need in order to supply the
information for the expert supply problem?”
Logan replied, “In collaboration with the data scientists, we have determined what information we
need to aggregate. It looks like we have the correct information: the Tickets DPQ serves the longterm view of all tickets raised and resolved, the User Maintenance DPQ provides daily snapshots for
all expert profiles, and the Survey DPQ provides a log of all survey results from customers.”
“Awesome,” said Addison. “Perhaps we should create a new DPQ named something like Experts Supply DPQ, which takes asynchronous inputs from those three DPQs? Its first product can be called
supply recommendations, which uses an ML model trained using data aggregated from DPQs in surveys, tickets, and maintenance domains. The Experts Supply DPQ will provide daily recommendations data, as new data becomes available about tickets, surveys and expert profiles. The overall
design looks like Figure 14-4.”
Figure 14-4. Implementing the Experts Supply DPQ
“OK, that looks perfectly reasonable,” said Dana. “The services are already done; we just have to make
sure the specific endpoints exist in each of the source DPQs, and implement the new Experts Supply
DPQ.”
“That’s right,” said Logan. “One thing we need to worry about, though—trend analysis depends on
reliable data. What happens if one of the feeder source systems returns incomplete information for a
chunk of time? Won’t that throw off the trend analysis?”
396
|
Chapter 14: Managing Analytical Data
“That’s correct—no data for a time period is better than incomplete data, which makes it seem like
there was less traffic than there was,” Dana said. “We can just exempt an empty day, as long as it
doesn’t happen much.”
“OK, Addison, you know what than means, right?” Logan said.
“Yes, I certainly do—an ADR that specifies complete information or none, and a fitness function to
make sure we get complete data.”
ADR: Ensure that Expert Supply DPQ Sources Supply an Entire Day’s Data or None
Context
The Expert Supply DPQ performs trend analysis over specified time periods. Incomplete
data for a particular day will skew trend results and should be avoided.
Decision
We will ensure that each data source for the Expert Supply DPQ receives complete snapshots for daily trends or no data for that day, allowing data scientists to exempt that day.
The contracts between source feeds and the Expert Supply DPQ should be loosely coupled to prevent brittleness.
Consequences
If too many days become exempt because of availability or other problems, accuracy of
trends will be negatively impacted.
Fitness functions:
Complete daily snapshot. Check timestamps on messages as they arrive. Given typical message volume, any gap of more than one minute indicates a gap in processing, marking
that day as exempt.
Consumer-driven contract fitness function for Ticket DPQ and Expert Supply DPQ. To ensure
that internal evolution of the Ticket Domain doesn’t break the Experts Supply DPQ.
Sysops Squad Saga: Data Mesh
|
397
CHAPTER 15
Build Your Own Trade-Off Analysis
Monday, June 10, 10:01
The conference room somehow seemed more brightly lit than it did on that fateful day
in September when the business sponsors of the Sysops Squad were about to pull the
plug on the entire support contract business line. People in the conference room were
chatting with each other before the meeting started, creating an energy not seen in the
conference room for a long, long time.
“Well,” said Bailey, the main business sponsor and head of the Sysops Squad ticketing
application, “I suppose we should get things started. As you know, the purpose of this
meeting is to discuss how the IT department was able to turn things around and repair what was
nine months ago a train wreck.”
“We call that a retrospective,” said Addison. “And it’s really useful for discovering how to do things
better in the future, and to also discuss things that seemed to work well.”
“So then, tell us, what worked really well? How did you turn this business line around from a technical standpoint?” asked Bailey.
“It really wasn’t one single thing,” said Austen, “but rather a combination of a lot of things. First of all,
we in IT learned a valuable lesson about looking at the business drivers as a way to address problems and create solutions. Before, we always used to focus only on the technical aspects of a problem, and as a result never saw the big picture.”
“That was one part of it,” said Dana, “but one of the things that turned things around for me and the
database team was starting to work together more with the application teams to solve problems.
You see, before, those of us on the database side of things did our own thing, and the application
development teams did their own thing. We never would have gotten to where we are now without
collaborating and working together to migrate the Sysops Squad application.”
399
“For me it was learning how to properly analyze trade-offs,” said Addison. “If it weren’t for Logan’s
guidance, insights, and knowledge, we wouldn’t be in the shape we’re in now. It was because of
Logan that we were able to justify our solutions from a business perspective.”
“About that,” said Bailey, “I think I speak for everyone here when I say that your initial business justifications were what prompted us to give you one last shot at repairing the mess we were in. That was
something we weren’t accustomed to, and, well, quite frankly it took us by surprise—in a good way.”
“OK,” said Parker, “so now that we all agree things seem to be going well, how do we keep this pace
going? How do we encourage other departments and divisions within the company from getting
into the same mess we were in before?”
“Discipline,” said Logan. “We continue our new habit of creating trade-off tables for all our decisions,
continue documenting and communicating our decisions through architecture decision records,
and continue collaborating with other teams on problems and solutions.”
“But isn’t that just adding a lot of extra process and procedures to the mix?” asked Morgan, head of
the marketing department.
“No,” said Logan. “That’s architecture. And as you can see, it works.”
Throughout this book, the unifying example illustrates how to generically perform
trade-off analysis in distributed architectures. However, generic solutions rarely exist
in architecture and, if they do, are generally incomplete for highly specific architec‐
tures and the unique problems they bring. Thus, we don’t think that the communica‐
tion analysis covered in Chapter 2 is exhaustive, but rather a starting point for you to
add more columns for the unique elements entangled with your problem space.
To that end, this chapter provides some advice on how to build your own trade-off
analysis, using many of the same techniques we used to derive the conclusions pre‐
sented in this book.
Our three-step process for modern trade-off analysis, which we introduced in Chap‐
ter 2 is as follows:
• Find what parts are entangled together.
• Analyze how they are coupled to one another.
• Assess trade-offs by determining the impact of change to interdependent
systems.
We discuss some techniques and considerations for each step next.
400
|
Chapter 15: Build Your Own Trade-Off Analysis
Finding Entangled Dimensions
An architect’s first step in this process is to discover what dimensions are entangled,
or braided, together. This is unique within a particular architecture but discoverable
by experienced developers, architects, operations folks, and other roles familiar with
the existing overall ecosystem and its capabilities and constraints.
Coupling
The first part of the analysis answers this question for an architect: how are parts
within an architecture coupled to one another? The software development world has
a wide variety of definitions of coupling, but we use the simplest, most intuitive ver‐
sion for this exercise: if someone changes X, will it possibly force Y to change?
In Chapter 2, we described the concept of the static coupling between architecture
quanta, which provides a comprehensive structural diagram of technical coupling.
No generic tool exists to build this because each architecture is unique. However,
within an organization, a development team can build a static coupling diagram,
either manually or via automation.
For example, to create a static coupling diagram for a microservice within an archi‐
tecture, an architect needs to gather the following details:
• Operating systems/container dependencies
• Dependencies delivered via transitive dependency management (frameworks,
libraries, etc.)
• Persistence dependencies on databases, search engines, cloud environments, etc.
• Architecture integration points required for the service to bootstrap itself
• Messaging infrastructure (such as a message broker) required to enable commu‐
nication to other quanta
The static coupling diagram does not consider other quanta whose only coupling
point is workflow communication with this quantum. For example, if an AssignTicket
Service cooperates with the ManageTicket within a workflow but has no other cou‐
pling points, they are statically independent (but dynamically coupled during the
actual workflow).
Teams that already have most of their environments built via automation can build
into that generative mechanism an extra capability to document the coupling points
as the system builds.
For this book, our goal was to measure the trade-offs in distributed architecture cou‐
pling and communication. To determine what became our three dimensions of
dynamical quantum coupling, we looked at hundreds of examples of distributed
Finding Entangled Dimensions
|
401
architectures (both microservices and others) to determine the common coupling
points. In other words, all the examples we looked at were sensitive to changes to the
dimensions of communication, consistency, and coordination.
This process highlights the importance of iterative design in architecture. No archi‐
tect is so brilliant that their first draft is always perfect. Building sample topologies for
workflows (much as we do in this book) allows an architect or team to build a matrix
view of trade-offs, allowing quicker and more thorough analysis than ad hoc
approaches.
Analyze Coupling Points
Once an architect or team has identified the coupling points they want to analyze, the
next step is to model the possible combinations in a lightweight way. Some of the
combinations may not be feasible, allowing the architect to skip modeling those com‐
binations. The goal of the analysis is to determine what forces the architect needs to
study—in other words, which forces require trade-off analysis? For example, for our
architecture quantum dynamic coupling analysis, we chose coupling, complexity,
responsiveness/availability, and scale/elasticity as our primary trade-off concerns, in
addition to analyzing the three forces of communication, consistency, and coordina‐
tion, as shown in the ratings table for the “Parallel Saga(aeo) Pattern” on page 346,
appearing again in Table 15-1.
Table 15-1. Ratings for the Parallel Saga pattern
Parallel Saga
Ratings
Communication
Consistency
Coordination
Coupling
Complexity
Responsiveness/availability
Scale/elasticity
Asynchronous
Eventual
Centralized
Low
Low
High
High
When building these ratings lists, we considered each design solution (our named
patterns) in isolation, combining them only at the end to see the differences, shown
in Table 15-2.
402
|
Chapter 15: Build Your Own Trade-Off Analysis
Trade-Offs
Table 15-2. Consolidated comparison of dynamic coupling patterns
Pattern
Coupling level Complexity Responsiveness/
availability
Scale/elasticity
Epic Saga
Phone Tag Saga
Fairy Tale Saga
Time Travel Saga
Fantasy Fiction Saga
Horror Story
Parallel Saga
Anthology Saga
Very high
High
High
Medium
High
Medium
Low
Very low
Very Low
Low
High
High
Low
Medium
High
Very high
Low
High
Very low
Low
High
Very high
Low
High
Low
Low
Medium
Medium
Low
Low
High
High
Once we had analyzed each pattern independently, we created a matrix to compare
the characteristics, leading to interesting observations. First, notice the direct inverse
correlation between coupling level and scale/elasticity: the more coupling present in
the pattern, the worse its scalability. This intuitively makes sense; the more services
involved in a workflow, the more difficult for an architect to design for scale.
Second, we made a similar observation around responsiveness/availability and cou‐
pling level, which is not quite as direct as the preceding correlation but also signifi‐
cant: higher coupling leads to less responsiveness and availability because the more
services involved in a workflow, the more likely the entire workflow will fail based on
a service failure.
This analysis technique exemplifies iterative architecture. No architect, regardless of
their cleverness, can instantly understand the nuances of a truly unique situation—
and these nuances constantly present themselves in complex architectures. Building a
matrix of possibilities informs the modeling exercises an architect might want to do
in order to study the implications of permutating one or more dimensions to see the
resulting effect.
Assess Trade-Offs
Once you have built a platform that allows iterative “what if ” scenarios, focus on the
fundamental trade-offs for a given situation. For example, we focused on synchro‐
nous versus asynchronous communication, a choice that creates a host of possibilities
and restrictions—everything in software architecture is a trade-off. Thus, choosing a
fundamental dimension like synchronicity first limits future choices. With that
Finding Entangled Dimensions
|
403
dimension now fixed, perform the same kind of iterative analysis on subsequent deci‐
sions encouraged or forced by the first. An architect team can iterate on this process
until they have solved the difficult decisions—in other words, decisions with entan‐
gled dimensions. What’s left is design.
Trade-Off Techniques
Over time, the authors have created a number of trade-off analyses and have built up
some advice on how to approach them.
Qualitative Versus Quantative Analysis
You may have noticed that virtually none of our trade-off tables are quantitative—
based on numbers—but are rather qualitative—measuring the quality of something
rather than the quantity, which is necessary because two architectures will always dif‐
fer enough to prevent true quantitative comparisons. However, using statistical analy‐
sis over a large data set allows reasonable qualitative analysis.
For example, when comparing the scalability of patterns, we looked at multiple differ‐
ent implementations of communication, consistency, and coordination combinations,
assessing scalability in each case, and allowing us to build the comparative scale
shown in Table 15-2.
Similarly, architects within a particular organization can carry out the same exercise,
building a dimensional matrix of coupled concerns, and look at representative exam‐
ples (either within the existing organization or localized spikes to test theories).
We recommend you hone the skill of performing qualitative analysis, as few opportu‐
nities for true quantitative analysis exist in architecture.
MECE Lists
It is important for architects to be sure they are comparing the same things rather
than wildly different ones. For example, it’s not a valid comparison to compare a sim‐
ple message queue to an enterprise service bus, which contains a message queue but
dozens of other components as well.
A useful concept borrowed from the technology strategy world to help architects get
the correct match of things to compare is a MECE list, an acronym for mutually exclu‐
sive, collectively exhaustive:
Mutually exclusive
None of the capabilities can overlap between the compared items. As in the pre‐
ceding example, it is invalid to compare a message queue to an entire ESB
because they aren’t really the same category of thing. If you want to compare just
404
|
Chapter 15: Build Your Own Trade-Off Analysis
the messaging capabilities absent the other parts, that reduces the comparison to
two mutually comparable things.
Collectively exhaustive
This suggests that you’ve covered all the possibilities in the decision space, and
that you haven’t left out any obvious capabilities. For example, if a team of archi‐
tects is evaluating high-performance message queues and consider only an ESB
and simple message queue but not Kafka, they haven’t considered all the possibil‐
ities in the space.
The goal of a MECE list is to cover a category space completely, with no holes or
overlaps, as shown pictorially in Figure 15-1.
Figure 15-1. A MECE list is mutually exclusive and collectively exhaustive
The software development ecosystem constantly evolves, uncovering new capabilities
along the way. When making a decision with long-term implications, an architect
should make sure a new capability hasn’t just arrived that changes the criteria. Ensur‐
ing that comparison criteria is collectively exhaustive encourages that exploration.
The “Out-of-Context” Trap
When assessing trade-offs, architects must make sure to keep the decision in context;
otherwise, external factors will unduly affect their analysis. Often, a solution has
many beneficial aspects, but lacks critical capabilities that prevent success. Architects
need to make sure they balance the correct set of trade-offs, not all available ones.
For example, perhaps an architect is trying to decide whether to use a shared service
or shared library for common functionality within a distributed architecture, as illus‐
trated in Figure 15-2.
Trade-Off Techniques
|
405
Figure 15-2. Deciding between shared service or library in a distributed architecture
The architect facing this decision will begin to study the two possible solutions, both
via general characteristics discovered through research and via experimental data
from within their organization. The results of that discovery process lead to a tradeoff matrix such as the one shown in Figure 15-3.
Figure 15-3. Trade-off analysis for two solutions
406
|
Chapter 15: Build Your Own Trade-Off Analysis
The architect seems justified in choosing the shared library approach, as the matrix
clearly favors that solution…overall. However, this decision exemplifies the out-ofcontext problem—when the extra context for the problem becomes clear, the decision
criteria changes, as illustrated in Figure 15-4.
Figure 15-4. Shifting decision based on additional context
The architect continued to research not only the generic problem of service versus
library, but the actual context that applies in this situation. Remember, generic solu‐
tions are rarely useful in real-world architectures without applying additional
situation-specific context.
This process emphasizes two important observations. First, finding the best context
for a decision allows the architect to consider fewer options, greatly simplifying the
decision process. One common piece of advice from software sages is “embrace sim‐
ple designs,” without ever explaining how to achieve that goal. Finding the correct
narrow context for decisions allows architects to think about less, in many cases sim‐
plifying design.
Second, it’s critical for architects to understand the importance of iterative design in
architecture, diagramming sample architectural solutions to play qualitative “whatif ” games to see how architecture dimensions impact one another. Using iterative
design, architects can investigate possible solutions and discover the proper context
in which a decision belongs.
Trade-Off Techniques
|
407
Model Relevant Domain Cases
Architects shouldn’t make decisions in a vacuum, without relevant drivers that add
value to the specific solution. Adding those domain drivers back to the decision pro‐
cess can help the architect filter the available options and focus on the really impor‐
tant trade-offs.
For example, consider this decision by an architect as to whether to create a single
payment service or a separate service for each payment type, as illustrated in Figure
15-5.
Figure 15-5. Choosing between a single payment service or one per payment type
As we discussed in Chapter 7, architects can choose from a number of integrators and
disintegrators to assist this decision. However, those forces are generic—an architect
may add more nuance to the decision by modeling some likely scenarios.
For example, consider the first scenario, illustrated in Figure 15-6, to update a credit
card processing service.
In this scenario, having separate services provides better maintainability, testability,
and deployability, all based on quantum-level isolation of the services. However, the
downside of separate services is often duplicated code to prevent static quantum cou‐
pling between the services, which damages the benefit of having separate services.
408
|
Chapter 15: Build Your Own Trade-Off Analysis
Figure 15-6. Scenario 1: update credit card processing service
In the second scenario, the architect models what happens when the system adds a
new payment type, as shown in Figure 15-7.
Figure 15-7. Scenario 2: adding a payment type
The architect adds a reward points payment type to see what impact it has on the
architecture characteristics of interest, highlighting extensibility as a benefit of sepa‐
rate services. So far, separate services look appealing.
However, as in many cases, more complex workflows highlight the difficult parts of
the architecture, as shown in the third scenario in Figure 15-8.
In this scenario, the architect starts gaining insight into the real trade-offs involved in
this decision. Utilizing separate services requires coordination for this workflow, best
handled by an orchestrater. However, as we discussed in Chapter 11, moving to an
orchestrator likely impacts performance negatively and makes data consistency more
of a challenge. The architect could avoid the orchestrator, but the workflow logic
must reside somewhere—remember, semantic coupling can only be increased via
implementation, never decreased.
Trade-Off Techniques
|
409
Figure 15-8. Scenario 3: using multiple types for payment
Having modeled these three scenarios, the architect realizes that the real trade-off
analysis comes down to which is more important: performance and data consistency
(a single payment service) or extensibility and agility (separate services).
Thinking about architecture problems in the generic and abstract gets an architect
only so far. As architecture generally evades generic solutions, it is important for
architects to build their skills in modeling relevant domain scenarios to home in on
better trade-off analysis and decisions.
Prefer Bottom Line over Overwhelming Evidence
It’s easy for architects to build up an enormous amount of information in pursuit of
learning all the facets of a particular trade-off analysis. Additionally, anyone who
learns something new generally wants to tell others about it, especially if they think
the other party will be interested. However, many of the technical details that archi‐
tects uncover are arcane to nontechnical stakeholders, and the amount of detail may
overwhelm their ability to add meaningful insight into the decision.
Rather than show all the information they have gathered, an architect should reduce
the trade-off analysis to a few key points, which are sometimes aggregates of individ‐
ual trade-offs.
Consider the common problem an architect might face in a microservices architec‐
ture about the choice of synchronous or asynchronous communication, illustrated in
Figure 15-9.
410
|
Chapter 15: Build Your Own Trade-Off Analysis
Figure 15-9. Deciding between communication types
The synchronous solution orchestrator makes synchronous REST calls to communi‐
cate with workflow collaborators, whereas the asynchronous solution uses message
queues to implement asynchronous communication.
After considering the generic factors that point to one versus the other, the architect
next thinks about specific domain scenarios of interest to nontechnical stakeholders.
To that end, the architect will build a trade-off table that resembles Table 15-3.
Trade-Offs
Table 15-3. Trade-offs between synchronous and asynchronous communication
for credit card processing
Synchronous
advantage
Synchronous
disadvantage
Asynchronous
advantage
Customer must wait for
credit card approval
process to start
No wait for process to
start
Credit approval is
guaranteed to start
before customer
request ends
Asynchronous
disadvantage
No guarantee that
the process has
started
Customer application
rejected if orchestrator is
down
Application
submission not
dependent on
orchestrator
After modeling these scenarios, the architect can create a bottom-line decision for the
stakeholders: which is more important, a guarantee that the credit approval process
starts immediately or responsiveness and fault-tolerance? Eliminating confusing techni‐
cal details allows the nontechnical domain stakeholders to focus on outcomes rather
than design decisions, which help avoids drowning them in a sea of details.
Trade-Off Techniques
|
411
Avoiding Snake Oil and Evangelism
One unfortunate side effect of enthusiasm for technology is evangelism, which should
be a luxury reserved for tech leads and developers but tends to get architects in
trouble.
Trouble comes because, when someone evangelizes a tool, technique, approach, or
anything else people build enthusiasm for, they start enhancing the good parts and
diminishing the bad parts. Unfortunately, in software architecture, the trade-offs
always eventually return to complicate things.
An architect should also be wary of any tool or technique that promises any shocking
new capabilities, which come and go on a regular basis. Always force evangelists for
the tool or technique to provide an honest assessment of the good and bad—nothing
in software architecture is all good—which allows a more balanced decision.
For example, consider an architect who has had success in the past with a particular
approach and becomes an evangelist for it, as illustrated in Figure 15-10.
Figure 15-10. An architect evangelist who thinks they have found a silver bullet
This architect has likely worked on problems in the past where extensibility was a key
driving architecture characteristic and believes that capability will always drive the
decision process. However, solutions in architecture rarely scale outside narrow con‐
fines of a particular problem space. On the other hand, anecdotal evidence is often
compelling. How do you get to the real trade-off hiding behind the knee-jerk
evangelism?
While experience is useful, scenario analysis is one of an architect’s most powerful
tools to allow iterative design without building whole systems. By modeling likely
scenarios, an architect can discover if a particular solution will, in fact, work well.
412
|
Chapter 15: Build Your Own Trade-Off Analysis
In the example shown in Figure 15-10, an existing system uses a single topic to
broadcast changes. The architect’s goal is to add bid history to the workflow—should
the team keep the existing publish-and-subscribe approach or move to point-to-point
messaging for each consumer?
To discover the trade-offs for this specific problem, the architect should model likely
domain scenarios using the two topologies. Adding bid history to the existing
publish-and-subscribe design appears in Figure 15-11.
Figure 15-11. Scenario 1: Adding bid history to the existing topic
While this solution works, it has issues. First, what if the teams need different con‐
tracts for each consumer? Building a single large contract that encompasses every‐
thing implements the “Stamp Coupling for Workflow Management” on page 378
anti-pattern; forcing each team to unify on a single contract creates an accidental
coupling point in the architecture—if one team changes its required information, all
the teams must coordinate on that change. Second, what about data security? Using a
single publish-and-subscribe topic, each consumer has access to all the data, which
can create both security problems and PII (Personally Identifiable Information, dis‐
cussed in Chapter 14) issues as well. Third, the architect should consider the opera‐
tional architecture characteristic differences between the different consumers. For
example, if the operations team wanted to monitor queue depth and use auto-scaling
for bid capture and bid tracking but not for the other two services, using a single topic
prevents that capability—the consumers are now operationally coupled together.
To mitigate these shortcomings, the architect should model the alternative solution to
see if it addresses the preceding problems (and doesn’t introduce new intractable
ones). The individual queue version appears in Figure 15-12.
Trade-Off Techniques
|
413
Figure 15-12. Using individual queues to capture bid information
Each part of this workflow (bid capture, bid tracking, bid analytics, and bid history)
utilizes its own message queues and addresses many of the preceding problems. First,
each consumer can have their own contract, decoupling the consumers from each
other. Second, security access and control of data resides within the contract between
the producer and each consumer, allowing differences in both information and rate
of change. Third, each queue can now be monitored and scaled independently.
Of course, by this point in the book, you should realize that the point-to-point based
system isn’t perfect either but offers a different set of trade-offs.
Once the architect has modeled both approaches, it seems that the differences boil
down to the choices shown in Table 15-4.
Trade-Offs
Table 15-4. Trade-offs between point-to-point versus publish-and-subscribe
messaging
414
Point-to-point
Publish-and-subscribe
Allows heterogeneous contracts
More granular security access and data control
Individual operational profiles per consumer
Extensibility (easy to add new consumers)
|
Chapter 15: Build Your Own Trade-Off Analysis
In the end, the architect should consult with interested parties (operations, enterprise
architects, business analysts, and so on) to determine which of these sets of trade-offs
is more important.
Sometimes an architect doesn’t choose to evangelize something but is rather coerced
into playing an opposite foil, particularly for something that has no clear advantage.
Technologies develop fans, sometimes fervent ones, who tend to downplay disadvan‐
tages and enhance upsides.
For example, recently a tech lead on a project tried to wrangle one of the authors into
an argument about monorepo versus trunk-based development. Both have good and
bad aspects, a classic software architecture decision. The tech lead was a fervent sup‐
porter of the monorepo approach, and tried to force the author to take the opposing
position—it’s not an argument if two sides don’t exist.
Instead, the architect pointed out that it was a trade-off, gently explaining that many
of the advantages touted by the tech lead required a level of discipline that had never
manifested within the team in the past, but will surely improve.
Rather than be forced into taking the opposing position, instead the architect forced a
real-world trade-off analysis, not based on generic solutions. The architect agreed to
try the Monorepo approach but also gather metrics to make sure that the negative
aspects of the solution didn’t manifest. For example, one of the damaging antipatterns they wanted to avoid was accidental coupling between two projects because
of repository proximity, so the architect and team built a series of fitness functions to
ensure that, while technically possible to create a coupling point, the fitness function
prevented it.
Don’t allow others to force you into evangelizing something—bring
it back to trade-offs.
We advise architects to avoid evangelizing and to try to become the objective arbiter
of trade-offs. An architect adds real value to an organization not by chasing silver
bullet after silver bullet but rather by honing their skills at analyzing the trade-offs as
they appear.
Trade-Off Techniques
|
415
Sysops Squad Saga: Epilogue
Monday, June 20, 16:55
“OK, I think I finally get it. We can’t really rely on generic advice for our architecture—it’s
too different from all the others. We have to do the hard work of trade-off analysis
constantly.”
“That’s correct. But it’s not a disadvantage—it’s an advantage. Once we all learn how to
isolate dimensions and perform trade-off analysis, we’re learning concrete things about
our architecture. Who cares about other, generic ones? If we can boil the number of
trade-offs for a problem down to a small enough number to actually model and test
them, we gain invaluable knowledge about our ecosystem. You know, structural engineers have
built a ton of math and other predictive tools, but building their stuff is difficult and expensive. Software is a lot…well, softer. I’ve always said that testing is the engineering rigor of software development.
While we don’t have the kind of math other engineers have, we can incrementally build and test our
solutions, allowing much more flexibility and leveraging the advantage of a more flexible medium.
Testing with objective outcomes allows our trade-off analyses to go from qualitative to quantitative
—from speculation to engineering. The more concrete facts we can learn about our unique ecosystem, the more precise our analysis can become.”
“Yeah, that makes sense. Want to go to the after-work gathering to celebrate the big turnaround?”
“Sure.”
416
|
Chapter 15: Build Your Own Trade-Off Analysis
APPENDIX A
Concept and Term References
In this book, we’ve made several references to terms or concepts that are explained in
detail in our previous book, Fundamentals of Software Architecture. The following is a
forward reference for those terms and concepts:
Cyclomatic complexity: Chapter 6, page 81
Component coupling: Chapter 7, page 92
Component cohesion: Chapter 7, page 93
Technical versus domain partitioning: Chapter 8, page 103
Layered architecture: Chapter 10, page 135
Service-based architecture: Chapter 13, page 163
Microservices architecture: Chapter 12, page 151
417
APPENDIX B
Architecture Decision Record References
Each Sysops Squad decision in this book was accompanied by a corresponding Archi‐
tecture Decision Record. We consolidated all the ADRs here for easy reference:
“ADR: A short noun phrase containing the architecture decision” on page 5
“ADR: Migrate Sysops Squad Application to a Distributed Architecture” on page 61
“ADR: Migration Using the Component-Based Decomposition Approach” on page 79
“ADR: Use of Document Database for Customer Survey” on page 184
“ADR: Consolidated Service for Ticket Assignment and Routing” on page 211
“ADR: Consolidated Service for Customer-Related Functionality” on page 215
“ADR: Using a Sidecar for Operational Coupling” on page 241
“ADR: Use of a Shared Library for Common Ticketing Database Logic” on page 247
“ADR: Single Table Ownership for Bounded Contexts” on page 280
“ADR: Survey Service Owns the Survey Table” on page 281
“ADR: Use of In-Memory Replicated Caching for Expert Profile Data” on page 297
“ADR: Use Orchestration for Primary Ticket Workflow” on page 321
“ADR: Loose Contract for Sysops Squad Expert Mobile Application” on page 380
419
APPENDIX C
Trade-Off References
The primary focus of this book is trade-off analysis; to that end, we created a number
of trade-off tables and figures in Part II to summarize trade-offs around a particular
architecture concern. This appendix summarizes all the trade-off tables and figures
for easy reference:
Figure 6-25, “Relational databases rated for various adoption characteristics” on page
163
Figure 6-26, “Key-value databases rated for various adoption characteristics” on page
166
Figure 6-27, “Document databases rated for various adoption characteristics” on page
168
Figure 6-28, “Column family databases rated for various adoption characteristics” on
page 170
Figure 6-30, “Graph databases rated for various adoption characteristics” on page 172
Figure 6-31, “New SQL databases rated for various adoption characteristics” on page
174
Figure 6-32, “Cloud native databases rated for various adoption characteristics” on
page 175
Figure 6-33, “Time-series databases rated for various adoption characteristics” on
page 177
Table 8-1, “Trade-offs for the code replication technique” on page 223
Table 8-2, “Trade-offs for the shared library technique” on page 227
Table 8-3, “Trade-offs for the shared service technique” on page 233
421
Table 8-4, “Trade-offs for the Sidecar pattern / service mesh technique” on page 239
Table 9-1, “Joint ownership table split technique trade-offs” on page 256
Table 9-2, “Joint ownership data-domain technique trade-offs” on page 258
Table 9-3, “Joint ownership delegate technique trade-offs” on page 261
Table 9-4, “Joint ownership service consolidation technique trade-offs” on page 262
Table 9-5, “Background synchronization pattern trade-offs” on page 272
Table 9-6, “Orchestrated request-based pattern trade-offs” on page 277
Table 9-7, “Event-based pattern trade-offs” on page 279
Table 10-1, “Trade-offs for the Interservice Communication data access pattern” on
page 286
Table 10-2, “Trade-offs for the Column Schema Replication data access pattern” on
page 288
Table 10-3, “Trade-offs associated with the replicated caching data access pattern” on
page 293
Table 10-4, “Trade-offs associated with the data domain data access pattern” on page
295
Table 11-1, “Trade-offs for orchestration” on page 306
Table 11-2, “Trade-offs for the Front Controller pattern” on page 312
Table 11-3, “Stateless choreography trade-offs” on page 313
Table 11-4, “Stamp coupling trade-offs” on page 314
Table 11-5, “Trade-offs for the choreography communication style” on page 315
Table 11-6, “Trade-off between orchestration and choreography for ticket workflow”
on page 319
Table 11-7, “Updated trade-offs between orchestration and choreography for ticket
workflow” on page 320
Table 11-8, “Final trade-offs between orchestration and choreography for ticket
workflow” on page 320
Table 12-11, “Trade-offs associated with state management rather than atomic dis‐
tributed transactions with compensating updates” on page 356
Table 12-12, “Trade-offs associated with atomic distributed transactions and compen‐
sating updates” on page 363
Table 13-1, “Trade-offs for strict contracts” on page 371
422
|
Appendix C: Trade-Off References
Table 13-2, “Trade-offs for loose contracts” on page 372
Table 13-3, “Trade-offs for consumer-driven contracts” on page 375
Table 13-4, “Trade-offs for stamp coupling” on page 379
Table 14-1, “Trade-offs for the Data Warehouse pattern” on page 385
Table 14-2, “Trade-offs for the Data Lake pattern” on page 388
Table 14-3, “Trade-offs for the Data Mesh pattern” on page 394
Table 15-2, “Consolidated comparison of dynamic coupling patterns” on page 403
Table 15-3, “Trade-offs between synchronous and asynchronous communication for
credit card processing” on page 411
Table 15-4, “Trade-offs between point-to-point versus publish-and-subscribe messag‐
ing” on page 414
Trade-Off References
|
423
Index
A
abstractness of codebase, 67
distance from main sequence, 69
ACID (atomicity, consistency, isolation, dura‐
bility) transactions
about, 263-267
distributed transactions versus, 265
example, 264
BASE transactions versus, 267
database types
cloud native databases, 176
document databases, 169
graph databases, 173
NewSQL databases, 173
relational databases, 163, 164
time series databases, 178
afferent coupling, 66, 68
aggregate orientation, 165
GraphQL read-only aggregated data, 368
key-value databases, 166
NoSQL databases understanding, 171
agility
modularity providing, 49, 49
speed-to-market via, 49
testability for, 54
versioning shared libraries, 225
Amazon DynamoDB key-value database, 165
Ambler, Scott, 154
analytical data
about, 4, 381
data meshes, 389-394
analytical reporting coupled, 393
when to use, 393
definition, 5
domain over technical partitioning, 388
previous approaches
data lakes, 386-389
data warehouses, 382-385
Sysops Squad saga, 381, 386, 389
data mesh, 394-397
annotations
distributed transaction management, 356
metadata for, 222
Anthology Saga pattern, 349-351
Apache Cassandra column family database, 170
Apache Ignite replicated caching, 290
APIs
endpoint versioning for shared services, 230
platforms for code reuse, 244
Architectural Decision Records (ADRs), 5
architectural decomposition
about, 64
eating the elephant, 63
codebase analysis
about, 65
abstractness and instability, 67
afferent and efferent coupling, 66
distance from main sequence, 69
component-based, 71-72
about, 64, 82
about patterns, 71, 82
architecture stories, 84
create component domains, 120-122
create domain services, 126-129
determine component dependencies,
111-118
flatten components, 101-107
425
gather common domain components,
94-96
identify and size components, 84-93
services built from components, 71
data
about, 132
about decomposing monolithic, 151
assign tables to data domains, 156-158
create data domains, 156
drivers of decomposition, 132-149
Refactoring Databases (Ambler and
Sadalage), 154
schemas to separate servers, 159
separate connections to data domains,
158
switch over to independent servers, 161
Sysops Squad saga, 150
tactical forking, 72-78
about, 64
trade-offs, 77
volatility-based decomposition, 191
architectural fitness functions (see fitness func‐
tions for governance)
architecture
architecture governance, 6
Big Ball of Mud anti-pattern, 65
Equifax data breach, 12
fitness functions, 7
(see also fitness functions)
architecture quantum, 28
(see also architecture quantum)
brittleness, 243
data warehouses, 384
strict contracts creating, 367
composition versus inheritance, 228
data importance, 4
(see also data)
definitions of terms, 14
architecture definition, 3
design versus, 13-15
security controlled through design, 215
evolution of, 3, 6, 47
(see also modularity)
iterative nature of analysis, 403
why more important than how, 13
architecture governance, 6
Big Ball of Mud anti-pattern, 65
Equifax data breach, 12
fitness functions, 7
426
|
Index
(see also fitness functions)
architecture quantum
about, 28, 144
coupling, static versus dynamic, 28
dynamic quantum coupling, 38-41
high static coupling, 30-37
user interface, 36
data disintegration driver, 144
separate databases, 159
Sysops Squad saga, 151
data product quantum, 390-393
high functional cohesion, 30
database, 145
independently deployable, 29
Sysops Squad saga, 42
architecture stories, 84
ArchUnit tool, 129
Aslett, Matthew, 173
asynchronous communication
about complexity of, 341, 345
definition, 14
delegate technique of data ownership, 259
distributed data access, 287
fault tolerance, 59
performance issues, 231
messaging mitigating, 231
publish-and-subscribe messaging, 277
dead letter queue, 279
durable subscribers, 278
synchronous versus, 38, 43
atomic fitness functions, 7
atomic workflow
consistency, 40
definition, 15
eventual consistency versus, 15, 40
atomicity in ACID transactions, 263
distributed transactions versus, 266
attributes for distributed transaction manage‐
ment, 356
availability
basic availability of BASE transactions, 267
coupling relationship, 403
database types
about availability, 162
cloud native databases, 176
column family databases, 170
document databases, 168
graph databases, 173
key-value databases, 166
NewSQL databases, 174
relational databases, 164
time series databases, 178
modularity and, 58
AWS Redshift cloud native database, 175
Azure CosmosDB cloud native database, 175
B
backpressure information link, 212
backup and restore databases, 160
bandwidth and stamp coupling, 377
BASE (basic availability, soft state, eventual
consistency) transactions, 267
Beck, Kent, 6
Berners-Lee, Tim, 4
best practices nonexistent, 1
Big Ball of Mud anti-pattern, 65
tactical forking, 72-78
bounded context in microservices, 4, 26, 135
architecture quanta, 35
architecture quantum versus, 42
breaking database changes controlled,
135-138
database abstraction, 136-138
data domains, 155
combining data domains, 157
data requirement in some architectures, 132
data sovereignty per service, 159
granularity and, 206
high functional cohesion, 30
ownership of data (see ownership of data)
breaking changes to data structure, 134
bounded context controlling, 135-138
database abstraction, 136-138
brittleness in architecture, 243
data warehouses, 384
strict contracts creating, 367
broker style event-driven architecture quan‐
tum, 33, 42
Brooks, Fred, 2
business case for modularity, 59-62
C
caching for distributed data access, 288-293
Chain of Responsibility design pattern, 338
The Checklist Manifesto (Gawande), 13
choreographed coordination, 306-311
about coordination, 40, 300
about orchestrated versus, 300, 331
trade-offs, 315
definition, 15
service granularity and, 200-203
stamp coupling for management, 378
workflow state management, 311-315
Front Controller pattern, 311
stamp coupling, 313
stateless choreography, 313
client/server applications, 382
cloud native databases, 175
Cockburn, Alistair, 235
CockroachDB NewSQL database, 174
code replication as reuse pattern, 221-223
code reuse patterns
about, 220
code replication, 221-223
granularity and shared code, 203-205, 224
orthogonal reuse pattern, 238
shared libraries
about, 223
granularity and, 224
versioning strategies, 225-227
shared services, 228-234
change risk, 229
fault tolerance, 232
performance, 231
scalability, 232
versioning via API endpoints, 230
sidecars and service mesh, 234-239
Sysops Squad saga, 239
strict contracts and, 367
Sysops Squad saga, 219
shared domain functionality, 244
sidecars and service mesh, 239
value added, 242
reuse via platforms, 244
code volatility as granularity driver, 191
cohesion
high functional cohesion, 30
database architecture quantum, 145
maintainability and, 51
metrics information link, 191
service granularity, 189, 194
column family databases, 169
communication
about synchronous versus asynchronous,
38, 43
analytical data mesh, 393
coordination required, 40
Index
|
427
database connection pool, 138
Sysops Squad saga, 150
distributed data access, 287
dynamic coupling as, 23, 29
communication dependencies, 29
saga pattern matrix, 41, 324, 402
east-west communication, 200-203
interservice communication, 200-203
API endpoint versioning for shared
services, 230
bounded context data, 206
data ownership, 252, 259
distributed data access, 285
performance issues, 231
latency with distributed data, 285
publish-and-subscribe messaging, 277
dead letter queue, 279
durable subscribers, 278
service scalability and elasticity, 57
compensating updates in distributed transac‐
tions, 275, 327
state management instead, 352, 355
Sysops Squad saga, 358-364
competitive advantage as modularity driver, 49
compile-based versus runtime changes, 229
component-based decomposition, 71-72
about, 64, 82
about components, 71
patterns
about, 71, 82
architecture stories, 84
create component domains, 120-122
create domain services, 126-129
determine component dependencies,
111-118
flatten components, 101-107
gather common domain components,
94-96
identify and size components, 84-93
components
Create Component Domains pattern,
120-122
fitness functions, 122
definition, 14, 51, 71, 103
dependencies determination, 112
Identify and Size Components pattern,
84-93
fitness functions, 87-89
maintainability and, 51
428
|
Index
names, 85
namespaces, 85
size determination, 84
size and granularity, 189
composite architecture characteristics, 7
composition versus inheritance in design, 228
connection management
connection quotas, 139
data disintegration driver, 138-141
scalability includes connections, 141
database per service, 143
Sysops Squad saga, 150
consistency
about, 40, 162
ACID transactions, 264
distributed transactions versus, 266
database types
cloud native ACID, 176
column family tunable, 171
document ACID, 169
graph ACID, 173
key-value tunable, 167
NewSQL ACID, 175
relational ACID, 164
time series ACID or tunable, 178
dynamic coupling relationship matrix, 40,
324, 402
eventual consistency
about patterns, 267
atomic workflow versus, 15, 40
background synchronization pattern,
269-272
BASE transactions, 267
event-based pattern, 277-279
finite state machines for, 351
orchestrated request-based pattern,
272-277
majority quorum, 167, 169
tunable consistency, 167, 171
Constantine, Larry, 66, 217
consumer-driven contracts, 373
context for trade-off analysis, 405-407
contracts
about, 365
bounded context data ownership, 136-138
contract fidelity and microservices, 373
consumer-driven contracts, 373
definition, 15
distributed data access
Data Domain not requiring, 294
Interservice Communication requiring,
286, 294
Replicated Cache requiring, 294
high static coupling, 30
loose contracts, 368
decoupling via, 369, 371
fitness functions required, 371
trade-offs, 371
microservice contracts, 372
stamp coupling
about, 376
bandwidth, 377
over-coupling via, 376
workflow management, 378
strict contracts, 367
JSON example, 367
trade-offs, 370
Sysops Squad saga, 365, 379
cooperative quantum, 393
coordination of workflow
about, 40, 300
about trade-offs, 315
choreographed coordination, 306-311
about orchestration versus, 300, 331
definition, 15
stamp coupling for management, 378
workflow state management, 311
dynamic coupling relationship matrix, 40,
324, 402
orchestrated coordination, 301-306
about choreographed versus, 300, 331
definition, 14
workflow state management, 311
stamp coupling for management, 378
Sysops Squad saga, 299, 317-321
coupling
about static versus dynamic, 29
dynamic coupling, 23, 29
static coupling, 23, 28
Sysops Squad saga, 42
architectural decomposition
afferent coupling, 66, 68
common domain functionality consoli‐
dated, 99
determining component dependencies,
112
efferent coupling, 66, 68
architecture quantum, 28, 34
dynamic quantum coupling, 38-41
high static coupling, 30-37
independently deployable, 29
Sysops Squad saga, 42
user interface, 36
availability relationship, 403
data relationships, 147
definition, 14, 27
interservice communication pattern, 286
loose contracts for decoupling, 369, 371
consumer-driven contracts, 373
microservices, 372
microservices and contract fidelity, 373
strict contract tight coupling, 370
maintainability and, 51
operational coupling, 234-239
Sysops Squad saga, 239
orchestration, 306
choreography versus, 332
orthogonal coupling, 238
scaling relationship, 332, 403
semantic coupling, 309
implementation coupling, 309
stamp coupling, 378
stamp coupling
bandwidth, 377
choreography state management, 313
over-coupling via, 376
workflow management, 378
trade-off analysis, 26-28, 401-403
coupling points, 402
static coupling diagram, 401
Create Component Domains pattern, 120-122
fitness functions, 122
Sysops Squad saga, 123-126, 129
Create Domain Services pattern, 126-129
fitness functions, 129
Sysops Squad saga, 129
credit card data (see PCI (Payment Card Indus‐
try) data)
custom attributes for distributed transaction
management, 356
cyclomatic complexity and maintainability, 51
D
data
about importance of, 4
about operational versus analytical, 4
analytical data
Index
|
429
about, 4, 381
data lakes, 386-389
data meshes, 389-394
data meshes coupled to reporting, 393
data warehouses, 382-385
definition, 5
domain over technical partitioning, 388
Sysops Squad saga, 381, 386, 389,
394-397
breaking changes, 134
bounded contexts controlling, 135-138
coupling
microservices, 35
multiple quanta event-driven architec‐
ture, 34
relationships among data, 147
data domains (see data domains)
database-per-service, 143
breaking apart database to achieve, 147
decomposition of operational data
about, 132
about decomposing monolithic, 151
assign tables to data domains, 156-158
create data domains, 156
data access in other domains, 158
database-per-service requiring, 147
dependencies, 153-155
drivers of, 132-149
Refactoring Databases (Ambler and
Sadalage), 154
schemas to separate servers, 159
separate connections to data domains,
158
switch over to independent servers, 161
Sysops Squad saga, 131, 150
disintegration drivers
about, 133
architecture quantum, 144
connection management, 138-141
database change control, 134-138
database type optimization, 146
fault tolerance, 143
scalability, 141
distributed data access (see distributed data
access)
granularity and data relationships, 205-207
integration drivers
about, 146
database transactions, 148
430
|
Index
relationships, 147
latency with distributed data, 285
ownership
about, 249
assigning, 250
common ownership, 252
data mesh domain ownership, 389
data sovereignty per service, 159
distributed data access, 287, 289
joint ownership, 253-260
service consolidation technique, 261
single ownership, 251
summary, 262
Sysops Squad saga, 249, 279-282
security
credit card data, 195
granularity driver, 195
services sharing, 132
changes to database, 134
connection management, 138-141
single point of failure, 143
single point of failure vulnerability, 143
Sysops Squad saga, 151
Sysops Squad data model, 19
data domains, 152
data domain technique of joint ownership,
256-258
data domains
about, 152
about decomposing monolithic data, 151
assign tables to data domains, 156-158
bounded context rules, 155
create data domains, 156
schemas to separate servers, 159
separate connections to data domains,
158
switch over to independent servers, 161
combining
service consolidation, 261
tables tightly coupled, 157
data access in other domains, 158
data schemas versus, 157
synonyms for tables, 157
distributed data access, 293-295
joint ownership of data, 256-258
soccer ball visualization, 153
Sysops Squad data model, 152
data lakes, 386-389
Data Mesh (Dehghani), 390
data meshes
about, 389
Data Mesh (Dehghani), 390
data product quantum, 390-393
Sysops Squad saga, 394-397
when to use, 393
data product quantum (DPQ), 390-393
aggregate DPQ, 391
cooperative quantum, 393
fit-for-purpose DPQ, 392
source-aligned (native) DPQ, 391
data warehouses, 382-385
Star Schema, 383, 384
technical partitioning, 384
database abstraction via bounded context,
136-138
Database as a Service (DBaaS), 164
NewSQL databases, 174
database refactoring book, 154
database structure Sysops Squad saga, 180-184
database transactions and granularity, 198
database types, 178
about characteristics rated, 161
aggregate orientation, 165
cloud native databases, 175
column family databases, 169
database type optimization, 146
document databases, 167
graph databases, 171-173
key-value databases, 165-167
NewSQL databases, 173
NoSQL databases, 165
relational databases, 163-164
schema-less databases, 169
Sysops Squad saga, 179
time series databases, 177-178
database-per-service, 143
breaking apart database to achieve, 147
databases separated physically, 159
Datomic cloud native database, 175
DDD (see Domain-Driven Design)
De La Torre, Fausto, 73
dead letter queue (DLQ), 279
decision making, 26-28
(see also trade-off analysis)
decomposition (see architectural decomposi‐
tion)
Decorator Design Pattern, 239
Dehghani, Zhamak, 389, 390
delegate technique of joint data ownership,
258-261
dependencies
Determine Component Dependencies pat‐
tern, 111-118
dynamic coupling as communication
dependencies, 29
monolithic data domains, 153-155
Refactoring Databases (Ambler and
Sadalage), 154
synonyms for tables, 157
shared libraries, 224, 227
static coupling as operational dependencies,
23, 28
tools
JDepend, 9, 66
NetArchTest, 11
visualization importance, 116
visualization tools resource, 113
deployment
architectural modularity described, 54
deployability in modularity, 55
fault tolerance, 58
cloud native Production Topology, 176
independent deployability, 29
deprecation strategies for shared libraries, 226
design
architecture versus, 13-15
security controlled through design, 215
Chain of Responsibility, 338
composition versus inheritance, 228
Decorator Design Pattern, 239
Pipes and Filters, 338, 351
Determine Component Dependencies pattern,
111-118
fitness functions, 117
Sysops Squad saga, 118-120
dimensional models of Star Schema, 384
distance from main sequence, 69
zones of uselessness and pain, 69
distributed architectures
bounded context in microservices, 26
(see also bounded context)
code reuse patterns, 220
(see also code reuse patterns)
data importance, 4
(see also data)
data sovereignty per service, 159
database connection pool, 138
Index
|
431
connection management, 139
Sysops Squad saga, 150
dynamic coupling matrix, 41, 324, 402
eventual consistency, 267
(see also eventual consistency)
evolution of architecture, 3
sagas as transactional behavior, 15
(see also sagas)
sharing data, 257
side effects, 361
state machines, 352-355
static coupling, 32
multiple quanta, 34
single quantum, 33
transactionality challenges, 40
distributed data access
Column Schema Replication pattern, 287
Data Domain pattern, 293-295
Interservice Communication pattern, 285
Replicated Caching pattern, 288-293
Sysops Squad saga, 283, 295-298
distributed transactions
about, 265, 327
ACID transactions, 263-267
example, 264
BASE transactions, 267
compensating updates, 275, 327
Sysops Squad saga, 358-364
eventual consistency, 267
(see also eventual consistency)
managing sagas, 356
(see also transactional saga patterns)
Sysops Squad sagas, 279-282, 358-364
distributed workflows
about, 299
about coordination, 40, 300
about trade-offs, 315
choreography, 306-311
about orchestration versus, 300
definition, 15
workflow state management, 311-315
orchestration, 301-306
about choreographed versus, 300
definition, 14
workflow state management, 311
semantic coupling, 309
implementation coupling, 309
service granularity and workflow, 200-203
stamp coupling for management, 378
432
|
Index
Sysops Squad saga, 299, 317-321
transactional saga patterns, 323
(see also transactional saga patterns)
document databases, 167
documentation for managing distributed trans‐
actions, 222, 356
domain partitioned architecture, 50
data domains, 152
data lakes losing relationships, 387
data mesh domain ownership, 389
granularity and shared domain functional‐
ity, 204
maintainability and, 51
technical partitioning
data warehouses as, 384
domain preferred, 388
domain workflows, 311
domain services
about, 126
component domains into, 121
Create Domain Services pattern, 126-129
fitness function for namespace, 129
domain cohesion and granularity, 194
Front Controller in choreography, 311
service-based architecture definition, 72,
126
domain versus infrastructure shared function‐
ality, 94
Domain-Driven Design (DDD)
bounded context, 4
architecture quantum, 42
high functional cohesion and, 30
information link, 121
namespaces from applications prior to, 121
durability in ACID transactions, 264
distributed transactions versus, 266
durable subscribers in pub/sub messaging, 278
dynamic coupling, 23, 29
about dimensions of, 300, 401
analytical data communication pattern, 393
communication, 38, 43
consistency, 40
coordination, 40
relationships among factors, 40, 324
(see also sagas)
static versus, 29
Sysops Squad saga, 42
east-west communication, 200-203
efferent coupling, 66, 68
elasticity
coupling, 43
definition, 56
mean time to startup and, 56
scalability versus, 43, 56
Enterprise Integration Patterns (Hohpe and
Woolf), 156
enterprise service bus (ESB) global orchestra‐
tor, 301
Epic Saga pattern, 325-330
Equifax data breach, 12
Evans, Erik, 165
event-based eventual consistency, 277-279
publish-and-subscribe messaging, 277
dead letter queue, 279
durable subscribers, 278
event-driven architecture (EDA)
multiple architecture quanta, 34
single architecture quantum
broker style as, 33, 42
mediator style as, 33
eventual consistency
about consistency, 40
about patterns, 267
atomic workflow versus, 15, 40
background synchronization pattern,
269-272
BASE transactions, 267
event-based pattern, 277-279
publish-and-subscribe messaging, 277
finite state machines for, 351
orchestrated request-based pattern, 272-277
extensibility as granularity driver, 196
eXtreme Programming (XP), 6
E
shared libraries, 232
shared services, 232
single point of failure, 143
Sysops Squad saga, 151
fitness functions for governance
about, 7-13
atomic, 7
composite architecture characteristics, 7
holistic, 8
manual versus automated, 11
outcomes, 11
overuse avoided, 13
dynamic coupling monitored, 29
examples
component cycles avoided, 8
layer relationship verification, 10
loose coupling requiring, 371
patterns for component-based decomposi‐
tion
create component domains, 122
create domain services, 129
determine component dependencies,
117
flatten components, 107
gather common domain components, 95
identify and size components, 87-89
zero-day exploits, 12
Flatten Components pattern, 101-107
fitness functions, 107
metrics for shared components, 106
Sysops Squad saga, 107-110
Foote, Brian, 65
Ford, Neal, xi, 417
Fowler, Martin, 165, 228
Front Controller in choreography, 311
Fundamentals of Software Architecture
(Richards and Ford), xi, 417-423
F
G
Fairy Tale Saga pattern, 333-336
Fantasy Fiction Saga pattern, 340-342
fault tolerance
choreography, 314
data disintegration driver, 143
definition, 58, 143, 193
granularity disintegration driver, 193
granularity integration driver, 200-203
modularity driver, 49, 58
orchestration, 305
Gather Common Domain Components pat‐
tern, 94-96
fitness functions, 95
Sysops Squad saga, 97-99
Gawande, Atul, 13
Gee, Trisha, 222
granularity
balancing trade-offs, 208
definition, 57
disintegrators of granularity
Index
|
433
about, 187, 188
code volatility, 191
extensibility, 196
fault tolerance, 193
scalability and throughput, 192
security, 195
service scope and function, 189
elasticity as function of, 57
integrators of granularity
about, 187, 197
data relationships, 205-207
database transactions, 198
shared code, 203-205
workflow and choreography, 200-203
metrics, 187
microservice single responsibility, 186, 190
modularity versus, 186
shared data and, 257
shared libraries, 224
Sysops Squad saga, 185, 209-216
graph databases, 171-173
GraphQL loose contracts, 368
Gremlin database language, 173
gRPC protocol
information link, 230
performance mitigation by, 231
strict contracts by default, 367
J
Java Message Service (Richards), 231
JDepend tool, 9
Eclipse plug-in, 66
joint ownership of data
about, 253
data domain technique, 256-258
delegate technique, 258-261
service consolidation technique, 261
table split technique, 254-256
K
H
hard parts of architecture, 2
hash table data structure, 165
Hello2morrow, 50
Henney, Kevlin, 222
Hexagonal Architecture, 235
high functional cohesion, 30
database architecture quantum, 145
Hohpe, Gregor, 156, 364
Horror Story pattern, 343-345
I
Identify and Size Components pattern, 84-93
fitness functions, 87-89
Sysops Squad saga, 81
implementation coupling, 309
independent deployability, 29
InfluxDB time series database, 177
infrastructure versus domain shared function‐
ality, 94
inheritance versus composition in design, 228
434
instability of codebase, 68
distance from main sequence, 69
interservice communication, 200-203
API endpoint versioning for shared services,
230
bounded context data, 206
data ownership
common ownership, 252
delegate technique of joint ownership,
259
distributed data access, 285
isolation in ACID transactions, 264
distributed transactions versus, 266
importance of, 361
isomorphic diagrams, 325
|
Index
key-value databases, 165-167
Kx time series database, 177
L
latency in communication, 285
LATEST caution for versioning, 227
layered architecture
architecture governance, 10
modular architecture
maintainability, 51
scalability and elasticity, 57
testability, 54
loose contracts
decoupling via, 369
fitness functions required, 371
REST and GraphQL, 368
trade-offs, 371
M
maintainability of modular architecture, 50-54
metric for tracking, 50
majority quorum, 167, 169
Martin, Robert C., 67, 190
mean time to startup and elasticity, 56
MECE (Mutually Exclusive, Collectively
Exhaustive) lists, 404
mediator in orchestration, 301
mediator style event-driven architecture quan‐
tum, 33
MemcacheDB key-value database, 165
messaging asynchronicity, 231
Java Message Service (Richards), 231
metadata annotations, 222
microkernel architecture, 50
microservices
architecture quantum, 28, 29
decoupled services, 35
high functional cohesion, 30
user interface coupling, 36
bounded context, 4, 26
architecture quanta, 35
architecture quantum versus, 42
breaking database changes controlled,
135-138
data domains, 155
data domains combined, 157
data requirement, 132
database abstraction, 136-138
granularity and, 206
high functional cohesion, 30
ownership of data (see ownership of
data)
contracts
about, 372
consumer-driven contracts, 373
contract fidelity, 373
coupling levels, 372
coordination, 40
(see also coordination)
definition, 191
domain partitioning of, 388
modularity, 52
deployability, 56
scalability and elasticity, 57
operational coupling, 234-239
orchestrator per workflow, 301
saga pattern, 323
service-based architecture as steppingstone, 72, 121, 127
shared database changed, 134
single responsibility principle and granular‐
ity, 186, 190
Microservices Patterns (Richardson), 323
Microsoft SQL Server relational database, 164
modularity
about, 48, 50
about change, 47
architectural governance, 8
drivers of
about, 49
agility, 49, 49
deployability, 55
fault tolerance, 58
maintainability, 50-54
scalability, 48, 56-57
testability, 54
granularity versus, 186
monolithic architectures
about modularity, 50
about need for modularity, 47
deployability, 55
fault tolerance, 58
maintainability, 51
modular monoliths, 50
scalability and elasticity, 57
Sysops Squad saga, 45, 59-62
testability, 54
water glass analogy, 47
monolithic architectures
architectural decomposition (see architec‐
tural decomposition)
Big Ball of Mud anti-pattern, 65
tactical forking, 72-78
data decomposition
about, 151
assign tables to data domains, 156-158
create data domains, 156
Refactoring Databases (Ambler and
Sadalage), 154
schemas to separate servers, 159
separate connections to data domains,
158
switch over to independent servers, 161
data importance, 4
database connection pool, 138
database type optimization, 146
Epic Saga mimicking, 325-330
modular architecture
Index
|
435
about need for, 47
deployability, 55
fault tolerance, 58
maintainability, 51
modular monoliths, 50
scalability and elasticity, 57
Sysops Squad saga, 45, 59-62
testability, 54
water glass analogy, 47
namespaces prior to Domain-Driven
Design, 121
patterns for migration to service-based
about, 71, 82
architecture stories, 84
create component domains, 120-122
create domain services, 126-129
determine component dependencies,
111-118
flatten components, 101-107
gather common domain components,
94-96
identify and size components, 84-93
single architecture quantum, 29
functional cohesion, 30
static coupling, 31
Monorepo, 415
MySQL relational database, 164
N
namespaces
component domains, 121
applications prior to Domain-Driven
Design, 121
fitness functions for governance, 122
components, 85
extend then prior a subdomain, 102
fitness function, 129
root namespace definition, 103
Neo4j graph database, 173
NetArchTest tool, 11
network latency, 285
NewSQL databases, 173
97 Things Every Java Programmer Should
Know (Henney and Gee), 222
NoSQL databases
aggregate orientation understood, 171
column family databases, 169
document databases, 167
key-value databases, 165
436
|
Index
NoSQL Distilled (Sadalage and Fowler), 165
schema-less databases, 169
Nygard, Michael, 5
O
online resources (see resources online)
Online Transactional Processing (OLTP), 4
operational coupling, 234-239
Sysops Squad saga, 239
operational data definition, 4
Oracle Coherence replicated caching, 290
Oracle relational database, 164
orchestrated coordination, 301-306
about choreographed versus, 300, 331
trade-offs, 315
about coordination, 40, 300
definition, 14
orchestrator, 301
workflow state management, 311
orchestrated request-based eventual consis‐
tency, 272-277
Orchestrated Saga pattern, 325-330
orphaned classes definition, 103
orthogonal code reuse pattern, 238
orthogonal coupling, 238
out-of-context trap, 405-407
ownership of data
about, 249
assigning, 250
common ownership, 252
data mesh domain ownership, 389
data sovereignty per service, 159
distributed data access, 287, 289
joint ownership
about, 253
data domain technique, 256-258
delegate technique, 258-261
table split technique, 254-256
service consolidation technique, 261
single ownership, 251
summary, 262
Sysops Squad saga, 249, 279-282
P
Page-Jones, Meilir, 23
Parallel Saga pattern, 346-348
PCI (Payment Card Industry) data
security and granularity, 195
Sysops Squad saga, 212-216
Personally Identifiable Information (PII), 387
Phone Tag Saga pattern, 330-333
PII (Personally Identifiable Information), 387
Pipes and Filters design pattern, 338, 351
platforms for code reuse, 244
Ports and Adaptors Pattern, 235
PostgreSQL relational database, 164
publish-and-subscribe messaging, 277
dead letter queue, 279
durable subscribers, 278
pull model consumer-driven contracts, 373
push model definition, 373
Q
quanta, 28
(see also architecture quantum)
quorum mechanism, 167, 169
R
Redis key-value database, 165
Refactoring Databases (Ambler and Sadalage),
154
relational databases (RDBMS), 163-164
links to most popular, 164
transactions and service granularity, 198
replicate databases, 160
replicated in-memory caching, 288-293
resources online
ArchUnit tool, 129
book website, xii
cohesion metrics information, 191
database links
cloud native databases, 175
NoSQL databases, 165
relational databases, 164
dependency tools
JDepend, 9, 66
NetArchTest, 11
visualization, 113
Refactoring Databases (Ambler and Sada‐
lage), 154
Saga Pattern (Richardson), 323
stamp coupling information, 376
static code analysis tools, 85
REST (representational state transfer) loose
contracts, 368
reuse patterns (see code reuse patterns)
Riak KV key-value database, 165
Richards, Mark, xi, 231, 417
Richardson, Chris, 323
root namespace definition, 103
runtime versus compile-based changes, 229
S
Sadalage, Pramod, 154, 165
sagas
about, 323
definition, 15
dynamic coupling matrix, 41, 324, 402
microservice saga pattern, 323
state machines, 352-355
Sysops Squad saga background, 15
(see also Sysops Squad sagas)
techniques for managing, 356
transactional saga patterns
about, 41, 324
about state management, 351
Anthology Saga, 349-351
Epic Saga, 325-330
Fairy Tale Saga, 333-336
Fantasy Fiction Saga, 340-342
Horror Story, 343-345
Parallel Saga, 346-348
Phone Tag Saga, 330-333
Sysops Squad saga, 323, 358-364
Time Travel Saga, 336-339
scalability
choreography, 314, 331
connection management, 141
database-per-service, 143
Sysops Squad saga, 150
coupling relationship, 332, 403
data disintegration driver, 141
database types
cloud native scalable, 176
column family scalable, 170
document scalable, 168
graph read versus write, 172
key-value scalable, 166
NewSQL scalable, 174
relational complexity, 164
time series scalable, 178
definition, 56
elasticity versus, 43, 56
granularity disintegration driver, 192
modularity providing, 47, 57
modularity driver, 49, 56-57
orchestration, 305
Index
|
437
choreography versus, 331
service consolidation for data ownership,
262
shared services, 232
schemas for database tables, 156
cross-schema access resolved, 158
data domain versus data schema, 157
databases separated physically, 159
schema-less databases, 169
synonyms for tables, 157
Scylla column family database, 171
security
granularity disintegration driver, 195
PCI data Sysops Squad saga, 212-216
security latency, 285
semantic coupling, 309
implementation coupling, 309
stamp coupling to manage workflow, 378
service consolidation technique of joint owner‐
ship, 261
service granularity (see granularity)
service mesh, 237-239
Sysops Squad saga, 239
service-based architecture
about, 32, 72, 126
data access in other domains, 158
data shared by services, 132
changes to database, 134
connection management, 138-141
single point of failure, 143
domain services from component domains,
121
Create Domain Services pattern, 126-129
evolution of architecture, 3
modular architecture scalability and elastic‐
ity, 57
monolithic migration to
about, 71, 82
architecture stories, 84
create component domains, 120-122
create domain services, 126-129
determine component dependencies,
111-118
flatten components pattern, 101-107
gather common domain components,
94-96
identify and size components, 84-93
static coupling, 32
438
|
Index
stepping-stone to microservices, 72, 121,
127
services
bounded contexts (see bounded context)
choreographed coordination definition, 15
component-based decomposition, 71
Create Component Domains pattern,
120
coupling definition, 14
(see also coupling)
data access in other domains, 158
data shared by services, 132
changes to database, 134
connection management, 138-141
single point of failure, 143
database connection pool, 138
Sysops Squad saga, 150
database-per-service, 143
definition, 14
distributed data access
Column Schema Replication pattern,
287
Data Domain pattern, 293-295
Interservice Communication Pattern,
285
Replicated Caching pattern, 288-293
Sysops Squad saga, 283, 295-298
extensibility, 196
granularity (see granularity)
identifying in sagas, 357
modularity
deployability, 56
scalability and elasticity, 57
testability, 54
orchestrated coordination definition, 14
ownership of data
assigning, 250
common ownership, 252
data sovereignty per service, 159
distributed data access, 287, 289
joint ownership, 253-260
service consolidation technique, 261
single ownership, 251
summary, 262
Sysops Squad saga, 249, 279-282
service mesh, 237-239
shared services for code reuse, 228-234
Sidecar pattern, 234-239
sharding in databases, 167
document databases, 168
NewSQL databases, 174
shared code and granularity, 203-205
(see also code reuse patterns)
shared libraries, 224
shared domain functionality
granularity and, 204
shared infrastructure functionality versus,
94
Sysops Squad saga, 244
shared libraries
about, 223
fault tolerance, 232
granularity and, 224
versioning strategies, 225-227
shared services for code reuse, 228-234
change risk, 229
fault tolerance, 232
performance, 231
scalability, 232
Sysops Squad saga, 244
versioning via API endpoints, 230
side effects in distributed architectures, 361
Sidecar pattern, 234-239
Sysops Squad saga, 239
silver-bullet solutions, 2
single point of failure (SPOF), 143
Sysops Squad saga, 151
single responsibility principle, 190
granularity and, 186, 190
SMS information link, 189
Snowflake cloud native database, 175
soft state of BASE transactions, 267
software architecture (see architecture)
SOLID principles (Martin), 190
SonarQube code hygiene tool, 10
speed-to-market as modularity driver, 49
stamp coupling
about, 376
bandwidth, 377
choreography state management, 313, 332
over-coupling via, 376
workflow management, 378
standard deviation calculation, 88
Star Schema of data warehouses, 383, 384
Starbucks Does Not Use Two-Phase Commit
(Hohpe), 364
state machines, 352-355
state management
compensating updates versus, 355
finite state machines for, 351
about state machines, 352-355
orchestration, 305
stamp coupling for workflow management,
378
workflow state management, 311-315
stateless choreography, 313
statements counted, 84, 187
static code analysis tools resource, 85
static coupling, 23, 28
architecture quantum, 34
high static coupling, 30-37
independently deployable, 29
Sysops Squad saga, 42
dynamic versus, 29
Sysops Squad saga, 42
operational dependencies as, 23, 28
trade-off analysis, 401
coupling points, 402
static coupling diagram, 401
Stine, Matt, 56
strict contracts, 367
JSON example, 367
trade-offs, 370
Structured Design (Yourdon and Constantine),
66
survey database structure, 180-184
synchronous communication
asynchronous versus, 38, 43
definition, 14
delegate technique of data ownership, 259
fault tolerance, 59
granularity and, 200-203
scalability and elasticity versus, 57
synonyms for tables, 157
Sysops Squad sagas
about, 15
nonticketing workflow, 16
problem, 17
ticketing workflow, 17
analytical data, 381, 386, 389
data mesh, 394-397
architectural components, 18
architecture quantum, 42
contracts, 365, 379
data model, 19
ACID transaction example, 264
data domains, 152
Index
|
439
distributed data access, 283, 295-298
Column Schema Replication pattern,
287
Data Domain pattern, 293-295
Interservice Communication Pattern,
285
Replicated Caching pattern, 288-293
distributed transactions, 279-282
distributed workflows, 299, 317-321
epilogue, 416
eventual consistency
about, 267
background synchronization, 269-272
event-based pattern, 277-279
orchestrated request-based pattern,
272-277
monolithic application broken apart, 45
6. create domain services, 127
business case for, 59-62
code reuse patterns, 219, 239, 244
create component domains, 123-126
create domain services, 129
data pulled apart, 131, 150
database structure, 180-184
database types, 179
decomposition method, 63, 78
decomposition pattern, 81
determine component dependencies,
118-120
flatten components, 107-110
gather common domain components,
97-99
granularity, 185, 209-216
identify and size components, 90-93
ownership of data, 249, 279-282
state machines, 352-355
trade-off analysis, 399, 416
transactional sagas, 323, 358-364
T
table split technique of joint ownership,
254-256
tactical forking, 72-78
about, 64
trade-offs, 77
technical partitioning
data warehouses as, 384
domain partitioning preferred, 388
domain workflows, 311
440
|
Index
maintainability and, 51
testability as modularity driver, 54
Thoughtworks
Architectural Decision Records, 5
composition versus inheritance, 228
time series databases, 177-178
Time Travel Saga pattern, 336-339
TimeScale time series database, 177
tools
ArchUnit, 129
dependencies
JDepend, 9, 66
NetArchTest, 11
visualization, 66, 113, 116
SonarQube code hygiene tool, 10
static code analysis, 85
visualization
about importance of, 116
afferent and efferent coupling, 66
data domains via soccer ball, 153
dependencies, 66, 113, 116
distance from main sequence, 69
domain diagramming exercise, 123
graph databases, 171-173
isomorphic diagrams, 325
static coupling diagram, 401
trade-off analysis, 26-28
assessing trade-offs, 403
least worst combination, 2
building your own, 400-403
consistency and availability, 178
coordination, 315
coupling, 26-28, 401-403
coupling points, 402
dynamic quantum coupling factors, 41
static coupling diagram, 401
documenting versus governing, 5
granularity, 208
iterative nature of, 403
loose contracts, 371
strict contracts, 370
Sysops Squad saga, 399, 416
tactical forking, 77
techniques of
bottom line decision, 410
MECE lists, 404
modeling relevant domain cases,
408-410
out-of-context trap, 405-407
qualitative versus quantitative analysis,
404
snake oil and evangelism, 412
transactional saga patterns
about, 324
pattern matrix, 41, 324, 402
about sagas, 323
management techniques, 356
about state management, 351
about state machines, 352-355
compensating updates versus, 355
Anthology Saga, 349-351
Epic Saga, 325-330
Fairy Tale Saga, 333-336
Fantasy Fiction Saga, 340-342
Horror Story, 343-345
Parallel Saga, 346-348
Phone Tag Saga, 330-333
Sysops Squad saga, 323, 358-364
Time Travel Saga, 336-339
transactionality
distributed architecture challenge, 40
security versus, 212-216
transactions
data integration driver, 148
database transactions and granularity, 198
distributed transactions
about, 265, 327
ACID transactions, 263-267
BASE transactions, 267
compensating updates, 275, 327
eventual consistency, 267
(see also eventual consistency)
Sysops Squad saga, 279-282, 358-364
Trunk-based Development, 415
tunable consistency, 167, 171
U
user interface quantum, 36
V
version control systems
code change frequency metric, 191
LATEST caution, 227
shared code libraries, 225-227
granularity and, 204, 205
shared service API endpoint versioning, 230
strict contracts requiring, 370
visualization tools
afferent and efferent coupling, 66
data domains via soccer ball, 153
dependencies, 113
importance of visualization, 116
JDepend Eclipse plug-in, 66
distance from main sequence, 69
domain diagramming exercise, 123
graph databases, 171-173
isomorphic diagrams, 325
static coupling diagram, 401
volatility-based decomposition, 191
W
What Every Programmer Should Know About
Object-Oriented Design (Page-Jones), 23
why more important than how, 13
wide column databases, 169
Woolf, Bobby, 156
workflow and service granularity, 200-203
(see also distributed workflows)
Y
Yourdon, Edward, 66
Z
zero-day exploits, 12
Zitzewitz, Alexander von, 50
zones of uselessness and pain, 69
Index
|
441
About the Authors
Neal Ford is a director, software architect, and meme wrangler at Thoughtworks, a
software company and a community of passionate, purpose-led individuals who
think disruptively to deliver technology that addresses the toughest challenges, all
while seeking to revolutionize the IT industry and create positive social change. He’s
an internationally recognized expert on software development and delivery, especially
in the intersection of Agile engineering techniques and software architecture. Neal
has authored seven books (and counting), a number of magazine articles, and dozens
of video presentations and spoken at hundreds of developers conferences worldwide.
His topics include software architecture, continuous delivery, functional program‐
ming, cutting-edge software innovations, and a business-focused book and video on
improving technical presentations. Check out his website, Nealford.com.
Mark Richards is an experienced, hands-on software architect involved in the archi‐
tecture, design, and implementation of microservices architectures, service-oriented
architectures, and distributed systems in a variety of technologies. He has been in the
software industry since 1983 and has significant experience and expertise in applica‐
tion, integration, and enterprise architecture. Mark is the author of numerous techni‐
cal books and videos, including the Fundamentals of Software Architecture, the
“Software Architecture Fundamentals” video series, and several books and videos on
microservices as well as enterprise messaging. Mark is also a conference speaker and
trainer and has spoken at hundreds of conferences and user groups around the world
on a variety of enterprise-related technical topics.
Pramod Sadalage is director of data and DevOps at Thoughtworks. His expertise
includes application development, Agile database development, evolutionary data‐
base design, algorithm design, and database administration.
Zhamak Dehghani is director of emerging technologies at Thoughtworks. Previ‐
ously, she worked at Silverbrook Research as a principal software engineer, and Fox
Technology as a senior software engineer.
Colophon
The animal on the cover of Software Architecture: The Hard Parts is a black-rumped
golden flameback woodpecker (Dinopium benghalense), a striking species of wood‐
pecker found throughout the plains, foothills, forests, and urban areas of the Indian
subcontinent.
This bird’s golden back is set atop a black shoulder and tail, the reason for its pyroinspired name. Adults have red crowns with black-and-white spotted heads and
breasts, with a black stripe running from their eyes to the back of their heads. Like
other common, small-billed woodpeckers, the black-rumped golden flameback has a
straight pointed bill, a stiff tail to provide support against tree trunks, and four-toed
feet—two toes pointing forward and two backward. As if its markings weren’t distinc‐
tive enough, the black-rumped golden flameback woodpecker is often detected by its
call of “ki-ki-ki-ki-ki,” which steadily increases in pace.
This woodpecker feeds on insects, such as red ant and beetle larvae, underneath tree
bark using its pointed bill and long tongue. They have been observed visiting termite
mounds and even feeding on the nectar of flowers. The golden flameback also adapts
well to urban habitats, subsisting on readily available fallen fruit and food scraps.
Considered relatively common in India, this bird’s current conservation status is lis‐
ted as being of “least concern.” Many of the animals on O’Reilly covers are endan‐
gered; all of them are important to the world.
The cover image is a color illustration by Karen Montgomery, based on a black and
white engraving from Shaw’s Zoology. The cover fonts are URW Typewriter and
Guardian Sans. The text fonts are Adobe Minion Pro and Myriad Pro; the heading
font is Adobe Myriad Condensed; and the code font is Dalton Maag’s Ubuntu Mono.
There’s much more
where this came from.
Experience books, videos, live online
training courses, and more from O’Reilly
and our 200+ partners—all in one place.
©2019 O’Reilly Media, Inc. O’Reilly is a registered trademark of O’Reilly Media, Inc. | 175
Learn more at oreilly.com/online-learning
Download